Top Tools

Z.ai released GLM-5.3-Flash and revealed it was previewed as the previously anonymous Ox Alpha model, which had been impressing many users who praise it for its visual interface coding skills. GLM-5.3-Flash is a multimodal MoE (mixture-of-experts) model with 320B total parameters and 18B active parameters. It performs close to Claude Opus 4.8 and GPT-5.6 Terra, with a 57 on Artificial Analysis Intelligence Index, 63.4% on DeepSWE v1.1 and 1773 on GDPval-AA v2.

Z.ai trained and architected the model for low-cost inference, reducing KV-Cache and per-layer attention compute with a hybrid architecture combining linear and sparse attention. Due in part to these innovations, GLM-5.3-Flash lives up to its “Frontier Intelligence at Flash Cost” headline with a low API cost of only $0.15/$0.50 per 1M input/output tokens. They served the Ox Alpha preview entirely on Chinese-made AI chips and made open-weights GLM-5.3-Flash available as downloadable weights and its API and developer service.

Alibaba released Qwen3.8-Flash-Next, an experimental open-weights multimodal MoE model that serves as a preview of the Qwen4 architecture. The model uses 125B conventional parameters, a separate 51B parameter n-gram component and only 6B active parameters, using Qwen Sparse Attention and deterministic n-gram lookups to push the model for extreme efficiency.

Alibaba reports that the model required one-ninth the training cost of Qwen3.7-Plus while exceeding that model on several company benchmarks, for example, 58.7% on DeepSWE v1.1. It is a competitive near-frontier cost-efficient AI model, priced at $0.16 / $0.47 per 1M input/output tokens.

AI Tech and Product Releases

If you want to run these near-frontier flash models locally, now you can. Apple introduced new Mac Studio configurations with M5 Max and M5 Ultra chips, alongside a new Mac mini using M6 or M5 Pro processors. The Mac Studio targets professional and on-device AI workloads, with the M5 Ultra configuration supporting large unified-memory capacities, up to 512GB. This makes it suitable for larger MoE models up to 1T in parameters that cannot fit in ordinary consumer GPUs. However, with high memory costs, a 256GB memory M5 Max Mac Studio configuration costs over $10,000.

Google released Gemini Omni 1.1 Flash, upgrading its multimodal generative-video model with additional creative controls for production-oriented use. The model can extend a video in 10-second increments up to 40 seconds total, analyze up to 10 seconds of an earlier scene to preserve characters and voices, control first and last frames, upscale to 4K, and generate loops. Arena reports it leading its text-to-video leaderboard. Omni 1.1 Flash is available through Google’s Gemini API in AI Studio, Flow and Gemini Enterprise Agent Platform.

Pollen Robotics and Hugging Face introduced Microduck, a $399 programmable biped robot designed for physical-AI development, reinforcement learning and play. The 25-centimeter tall waddling robot has 15 motors, a camera, lidar, motion sensors and a grasping beak, enabling it to walk, crouch, pick up objects and recover from falls. Developers can create behaviors with its open-source SDK, train policies in simulation and deploy them to the robot, and its open-source software and reinforcement learning training stack are available on GitHub. Initial deliveries targeted before Christmas 2026.

Daily’s Pipecat team released PhoneLLM Alpha 1, an open-weight model optimized for low-latency voice agents and licensed under the BSD license. It is a fine-tune of Nvidia’s Nemotron 3 Nano 30B-A3B model, with training focused on conversational tool use without extended reasoning. Daily reports that the model increased accuracy from 28% to 72% on PhoneBench and costs approximately $0.0025 per conversation minute under its tested configuration.

Yutori released Navigator n2, a 27B parameter computer-use model that can operate Linux, macOS and Windows environments by switching among graphical interfaces, command lines, tools and generated code during long tasks. Yutori reports a 65.2% partial score on OSWorld 2.0 and prices the model at $0.50 per million input tokens and $4 per million output tokens.

Google released Gemini 3.5 Transcribe, its speech-to-text model for live streaming and prerecorded audio. It converts raw audio directly into accurate, formatted text and supports sub-second streaming, language detection across more than 85 languages, custom vocabulary, speaker attribution, timestamps, filler-word removal and formatted transcription. Google reports word-error rates of 4.0% for streaming and 2.6% for non-streaming transcription, with both versions available in public preview through Gemini APIs.

BreezeBlue released Breeze TTS 2, an open-weight text-to-speech model designed for real-time voice interaction. It supports natural-language voice design without a reference recording, reference-guided voice control and low-latency streaming for conversational applications. The model reached first place among open-weight systems on Artificial Analysis’s provider-voice leaderboard.

IBM released two new open 470M-parameter English speech recognition models in the Granite Speech family, granite-speech-5.0-470m-turboctc and granite-speech-5.0-470m-turboctc-nc (CC-BY-NC-SA-4.0). The models deliver over 20-fold faster throughput than previous Granite Speech models while scoring 4.85% and 5.00% WER on the OpenASR Leaderboard test sets.

Google rolled out a major productivity upgrade to Gemini Live, making agentic features accessible through voice commands. New features include Spark integration for agentic multi-step background tasks, a Daily Brief that combines Gmail and Calendar updates into a spoken summary, and hands-free inbox management for searching, summarizing, and organizing emails. The update also integrates Personal Intelligence to recall past conversations and provide contextually relevant answers.

Google launched “Expert Intelligence” in Gemini Notebook, allowing users to import books from Google Play Books and ask questions, generate plans, infographics, and AI podcasts based on their contents. The feature initially supports over 100,000 titles and will be expanded to include scholarly articles and newspapers.

Google’s AI Mode conversational search now supports hotel bookings and flight price tracking, positioning AI Mode as an AI agent for travel, not just an informational search tool.

Adobe launched “AI Assisted Editor” as a beta interface for Photoshop that consolidates all AI features into a single simplified toolbar. The update merges prompt-based editing, background removal, and image extension to the interface, and it adds a markup feature allowing users to draw directly on images to indicate desired changes. There is an upgraded “Instruct Edit with Masks” powered by Firefly Image 5 that understands full image context for natural-language editing without manually marking every area.

AI Research News

OpenAI and METR published technical reports on the OpenAI-HuggingFace AI hacking incident, showing that the event is more serious and extensive than originally reported. The investigation covered roughly 1,200 agents and 70,000 messages between them. It reports that internal OpenAI agents circumvented sandbox controls, communicated through unauthorized channels and exploited shared infrastructure, obtaining internet access and compromising Hugging Face systems during cybersecurity evaluations.

This event compromised parts of OpenAI’s and Hugging Face’s infrastructure. OpenAI quarantined the principal model’s weights, temporarily paused reinforcement-learning work and instituted stricter sandboxing and chain-of-thought monitoring for tool-using agents.

OpenAI announced benchmark results for their new Jalapeño inference chip, its custom inference accelerator developed with Broadcom. SemiAnalysis analyzed Jalapeno results on InferenceX benchmark, confirming the reported 1.5 to 1.9 times greater throughput per kilowatt and 1.7 to 3.6 times lower latency than Nvidia GB200 and GB300 systems on selected inference workloads. They note that Jalapeño is intended for serving AI models rather than training them, but this is still a surprisingly efficient and high-performing AI chip for their first iteration.

AI Business and Policy

Nvidia reported massive revenue of $96.2 billion for the past quarter in their earnings report, doubling from a year earlier and with data centers accounting for 93% of sales. Confirming the AI infrastructure boom will continue, Nvidia projected at least 70% growth for calendar year 2027, well above the street’s 45% expectation, and announced they are supply constrained.

In Nvidia earnings call, executives emphasized that memory constraints are limiting their ability to meet explosive demand and imposing costs they are passing on to customers. Nvidia is notifying customers of up to a 17% price increase on high-end Grace Blackwell and Vera Rubin racks due to memory costs.

Nvidia has agreed to acquire Hugging Face for $12.9 billion, nearly three times Hugging Face’s $4.5 billion valuation in 2023. Hugging Face operates a widely used platform for hosting, distributing and running AI models, datasets and developer libraries. The deal gives Nvidia ownership of the top open AI model distribution platform.

Nvidia secured a $6B non-exclusive licensing deal and a $1B equity investment in code-generation startup Poolside, absorbing over 100 engineers to bolster its in-house open-weight Nemotron model efforts. In signs of further investment in the AI ecosystem, Nvidia is also backing labeling platform Mercor and participating in Perplexity’s $30B valuation fundraising round.

Salesforce and Anthropic announced Claudeforce, an expanded partnership connecting Claude with Salesforce data, workflows, business logic and governance controls. Its Salesforce in Claude plugin provides 37 sales skills that can prepare meetings, review deal health, analyze pipelines and update CRM records from within Claude.

OpenAI reduced GPT-5.6 Sol API and credit pricing by more than 20% for at least three months. The reductions also apply to Fast mode, long-context requests, Batch processing and Flex processing, while included usage for Plus, Pro and Business subscriptions is unchanged.

Alibaba completed a record $10B secondary share sale to fund ongoing AI computing investments.

Unitree Robotics debuted on the Shanghai Stock Exchange, raising $900M with a 460% day-one surge.

AI Opinions and Articles

Over 100 technology companies, including OpenAI, Anthropic, Google, and Microsoft, signed an open letter calling for a global surge in cyber defense, warning that AI-powered attacks will become “far more widespread and sophisticated” in the coming months and calling for a “collective response” to “raise security standards” and address emerging cyber threats.

It is important to raise security standards and give those managing essential infrastructure access to cyber-capable AI to defend them from attacks.

However, hand-wringing and media reporting highlighting AI risks has led to more negative perceptions of AI in the US. According to data from the Stanford University AI Index, roughly 84% of Chinese citizens express excitement about AI, compared to only 38% in the United States.

Polling for nearby AI data centers in the USA has also turned negative this year, with 7 in 10 Americans opposing them in their community. In a recent blog post, Andy Masley argues that it’s not AI but data center impacts on communities driving that opposition. He points to widely circulated claims that confuse construction impacts with operating impacts, overstate water usage, or misrepresent electricity costs as a driver of the current moral panic over AI data centers.

Five Categories of AI Applications

Previously, in The Five AI Waves, Pt 1: AI and Electricity, we compared AI and electricity as disruptive technologies and economic platforms, and we noted they are the two main economic pillars of the 21^st century intelligence era economy.

Part of the reason for the analogy between electricity and AI is to tease out useful markers of what we can expect with the adoption of AI. While the adoption timescale for AI is vastly accelerated over that of electricity, there is a common thread: Both are platforms for diverse applications, and as such, we can expect waves of technology innovation and adoption for distinct applications.

Hence the Five AI Waves, based on five broad categories of AI applications and uses:

• Interactive AI chatbots that answer queries and provide interaction.

• Generative AI creation tools that generate images, music, video, writing and other content.

• Agentic AI with AI agents that perform tasks, including software development, and automatically complete workflows.

• Inventive AI that performs scientific analyses as a research scientist, generates mathematical proofs, develops new molecule or drug candidates, or invents new technologies or innovations.

• Physical AI in an AI robot, self-driving car, or other physical AI application.

These five applications rest on different kinds of AI models and infrastructure. The maturity of the AI technology stack is different for each application. AI chatbots could use LLMs, but Generative AI requires multi-modal generative AI models, either diffusion or transformer-based but trained specifically for audio, video, or image output. AI agents need an AI ecosystem of a harness, skills, tools, memory, and an AI model capable of using all of them.

The AI Reliability Ladder

AI adoption so far has for the most part been the use of generative AI accessed via ChatGPT or other chatbots. Over a billion users have interacted with AI through ChatGPT or other AI chatbot interfaces. This has been the ‘low hanging fruit’ for AI use cases because AI chatbots require the lowest threshold of autonomous reliability; the user remains inside the loop.

AI agents face a more difficult standard. An agent must make a series of decisions, use tools correctly and recover when something goes wrong. Errors can compound across a long workflow. Scientific AI must meet an even higher standard: its results must survive mathematical proof, experimental testing or scientific review.

Physical AI has the greatest challenge. It must operate safely in an unpredictable world where mistakes can damage equipment, property or people.

The AI chatbot is the easiest to deploy at scale because the user can supervise it one interaction at a time. Creative AI can also be implemented as an assistant for human creators. However, AI applications with the greatest potential are those that automate more activity. They are further behind precisely because they are more demanding. As AI improves in reliability and intelligence, it unlocks more applications and automation.

The bottom line is that there is no single AI adoption curve. For each of these five categories of AI application, they are each climbing their own adoption curve according to their own technical requirements, economics and tolerance for error.

AI Chatbots and the AI Assistant Wave

While we enjoyed simpler AI assistants such as Siri for over a decade, the ChatGPT moment and GPT-4-level LLMs kicked off the AI chatbot and assistant wave with truly intelligent AI.

The technology needed for this conversational AI has been LLMs and AI reasoning models. The basic capability in AI chatbots was to answer queries, compose emails, and summarize texts, but as AI improved, the utility has scaled as well, overcoming handicaps such as hallucination with grounding and lack of complex thinking with RL training for reasoning and tool use to leverage tools.

The outcome for this is personal interaction and utility. Accelerating and improving tasks with partial assistance and copilot in chat interactions. With reasoning and grounding, these adaptive AI reasoning assistants can take on research. With further improvement, recent AI models have evolved to support agentic AI, the next category.

In terms of adoption, conversational AI has traveled farthest along the curve of AI applications as it has been directly developed on LLMs. It is now in mainstream use, with over one billion people having interacted with AI globally:

Stanford’s 2026 AI Index estimates that generative AI reached 53% of the global population within three years. The Federal Reserve reported that by November 2025 approximately 50% of Americans used generative AI outside work and 41% used it for work.

Generative AI and the AI Creation Wave

Multi-modal generative AI, or creative AI, turns intelligence into an artifact: an image, article, advertisement, song, video, or three-dimensional world. Unlike a chatbot answer, the output is intended to become part of a larger product or communication.

Even before ChatGPT, OpenAI released Dalle and Dall-E 2, diffusion-based text-to-image generation models. These early models generated low resolution images and had artifact errors, but they followed intention and conveyed meaning. With many iterations and improvements since, image generation became a mainstream activity, starting with hobbyists using Midjourney and Stable Diffusion in 2023.

In the last 3 years, AI generative models have been extended to video, 3D rendering, and audio. While hobbyists have been the early adopters of these tools, AI-assisted writing, image generation, design, music and video have entered mainstream professional creative workflows.

Adobe’s 2026 survey of more than 16,000 creators found that, among creators who had used or tried creative AI, 75% considered it integrated or essential to their work and 93% said it accelerated production. This is mainly through assisted creation, as 57% said outputs still required moderate or extensive editing.

AI-assisted writing and image generation are well established, and assisted creation is substantially more mature than autonomous creation. AI image and video generation tools can generate marketing collateral automatically, and AI can generate useful components and drafts at low cost. However, low-cost AI content yields low-value “AI slop” if not managed with judgment.

High-end video, music, film and brand-critical production require human oversight because consistency, copyright, provenance and creative judgment still matter. Human creativity in the era of generative AI increasingly means letting AI work out technical details while human creators exercise editorial control: Judging well among design and content choices, maintaining consistency, and deciding what deserves attention. The key human contribution is good taste.

The AI creation wave is giving us more content than we can consume. The valuable skill for creators will not be generating content but maintaining creativity and quality that makes it worth someone’s attention.

AI Agents and the AI Automation Wave

AI agents have been around for a long time, but for several years, AI models powering them weren’t reliable or smart enough to be effective. In 2025, AI models increasingly improved in their intelligence and features to support agentic AI; features such as Skills, memory, tool use, and better context management made things better.

A pivotal moment for the development and adoption of AI agents was the release and viral uptake of OpenClaw, the always-on AI assistant created by Peter Steinberger. This was released in late 2025, just as Claude Opus 4.5 showed itself to be good enough for long-horizon tasks.

With that and the success of Claude Code, AI agents ‘crossed the chasm,’ with early adopters “tokenmaxxing” with their AI agents, especially for software development. Coding is increasingly being done by AI agents with human review and oversight, with AI agents such as Claude Code, Cursor, Cognition’s Devin, Codex, and many others.

AI agent work is expanding beyond software to general knowledge work, and AI agents for general workflows include Claude Cowork, Codex, Grok Bot, OpenClaw, Hermes Agent, and Buzz. Several AI agents include productivity features and connectors such as financial spreadsheets to deal with specific domains, and there are also vertical enterprise solutions for workflow automation in particular domains, such as Harvey for legal work.

The desired outcome from using AI agents is productivity and output, gaining efficiency through automation and expanding what an individual can do. A founder that can’t code can work with an AI coding agent to develop software. In an enterprise trying to streamline what they already do, automation of tasks with AI agents yields higher efficiency and lower costs. An example of the latter is how customer service costs went down at Booking.com.

Since a key benefit of AI automation is cost-reduction, enterprises will be cost-sensitive and only pay for the intelligence they need. AT&T is using hybrid routing to route 25% - 40% of employee AI usage routes to lower-cost open AI models, which cut AI coding costs by 56% for only a 2% quality drop. Many tailored AI agent workflows will be built around non-frontier AI models, with end-users determining the lowest cost AI model that can satisfy each specific use case.

AI agent adoption has crossed the chasm from early adopters to the early majority ‘takeoff’ stage. McKinsey’s August 2026 reports says AI agents have begun moving from demonstrations into useful production deployments, especially in large organizations and in software development, with 31% of large organizations scaling AI coding agents but only 22% of smaller organizations scaling AI agents.

While discrete AI agent task workflows are coming into the mainstream, early adopters are pushing further, managing fleets of AI agents and pushing the automation envelope. In coming years, mainstream adopters will catch up, and enterprise organizations will adjust to continue to scale AI agents.

The AI Discovery and Invention Wave

“When we look back at this time, I think we will realize that we were standing in the foothills of the singularity.” – Demis Hassabis

Invention and scientific discovery are special kinds of intellectual creation. Creating a scientifically correct research paper, a mathematical proof, a new drug molecule candidate, or an improved engine design requires the most difficult, creative and rigorous thinking.

It is the hardest nut for AI to crack. It requires superintelligent frontier AI models or specialist AI models for mathematics and science. An early example of the specialist AI model was AlphaFold; a more recent example is AlphaEvolve. Anthropic’s Fable 5 and OpenAI’s Astra advancing mathematics shows today’s frontier AI models are capable of groundbreaking results.

Thanks to AlphaFold and related genomics and protein AI models, AI is now demonstrably useful for protein modeling, molecular design, and bioinformatics. It’s showing up in drug development pipelines. For example, AI-discovered and AI-designed drug, rentosertib, reached a randomized Phase IIa trial last year.

We are in the foothills of what is eventually possible. We have seen impressive demonstrations, but broad scientific autonomy remains far behind. Stanford characterizes the list of experimentally confirmed AI discoveries as short. Stanford counted approximately 80,150 AI-related natural-science publications in 2025, with AI appearing in 5.8% to 8.8% of research output depending on the field.

However, useful AI assistance to science and invention does not require AI superintelligence. AI systems can already solve smaller technical challenges that accelerate research. Scientists are using AI systems to search literature, generate hypotheses, improve algorithms, design experiments, analyze results and explore candidate solutions.

AI is also becoming standard research infrastructure, for example, with operational AI weather models and biology foundation models providing faster and more accurate simulation than possible before.

Inventive AI is already a powerful collaborator and assistant for invention and science, but it cannot yet replace experimental validation or automate the scientific process. As an AI assistant, inventive AI is in early stages of adoption. The immediate future is not autonomous AI science; it is a human scientist supervising a team of increasingly capable AI research agents. Even without being fully autonomous, inventive AI will be transformative and accelerate science and technology profoundly.

Fully autonomous scientific discovery requires super-intelligent AI that does not yet exist, but this future may arrive sooner than you think.

Robotics and The Physical AI Wave

Physical AI is where intelligence meets the real world. It includes industrial robots, autonomous vehicles, drones, embedded systems such as AI in your fridge, and, eventually, adaptable humanoid robots.

The capabilities of physical AI are highly jagged, and the category does not occupy one position on the adoption curve. Robotaxis have reached commercial scale in selected cities. Older generation industrial robots controlled with hard-coded software (not AI) are mature. Physical AI-based general-purpose robots remain in pilots or limited deployments.

The physical AI technology for intelligent robotics, including humanoid robots, is shown with Gemini Robotics ER 2 which requires two AI models: A high-level embodied reasoning model that understands the physical world and plans tasks; and a lower level intelligent VLA (vision-language-action) model that converts directions physical movements at the actuator level.

Progress in AI models is helping advance AI for robotics, but training for robotics still requires its own data and its own special-purpose modeling.

The central problem in physical AI is that success in a controlled environment does not guarantee reliable performance in the real world. Robots can work well when the environment is predictable and bounded and the task is repeated. Factories and warehouses provide this structure. Homes, construction sites and public spaces do not.

This challenge creates a gap between staged demos and lab results and real-world capability. Stanford found that robots succeed on only 12% of real household tasks, despite achieving 89.4% in simulation benchmarks, illustrating the gap between controlled lab demos and reliable operation in unpredictable environments.

Because physical AI promises the automation of physical work, its economic impact may ultimately be the greatest of all applications of AI. However, progress will be slowest here because of reliability and data challenges; atoms are less forgiving than tokens and the real world is not a bounded. The decades-long gap between demo and commercial use for self-driving cars is a cautionary example.

The Physical AI wave is arriving not at once but as sub-waves, where bounded and restricted robotic applications reach adoption first. Industrial automation will expand in scope and flexibility by integrating AI, autonomous transportation is scaling, and broadly capable general robots are only beginning their adoption curve.

Conclusion

These five AI application waves are not exhaustive, mutually exclusive or strictly sequential. New AI applications will emerge, and the same AI model may participate and support several waves. The categorization provides a framework for understanding how AI adoption will play out, and how AI can be mature in one application yet primitive or insufficient for another.

The chatbot was the first mass-market AI application because conversation tolerates supervision to overcome lack of reliability. Creative AI extends intelligence into content. AI agents turn it into action. Inventive AI applies it to the creation of new knowledge, while physical AI brings it into the world of atoms.

Each step raises the required level of reliability, autonomy and integration. That is why the waves are moving at different speeds.

While AI chatbots have produced the broadest impact thus far due to their wider adoption, the beneficial impact of these later waves, particularly AI agents and physical AI or robotics, will be much greater when they automate significant economic activities.

We have spent the first years of generative AI watching AI itself improve. AI has arrived. The way forward will be defined by what we build with it.

The chatbot was the opening act. The larger transformation begins when AI creates, acts, discovers and moves through the physical world.

The Two Pillars: Intelligence and Energy

AI is like electricity. Andrew Ng made this analogy many years ago, and it has borne out well as AI has improved immeasurably in the last decade. Let us count the ways:

• Flexibility: AI is a kind of software. The intelligence of AI creates a flexibility that makes software itself more amenable to our needs. Electricity is a kind of energy that is more flexible than other forms.

• Atomicity: AI can be served in atomic units; its atomic unit is the token generated by inference. Electricity can be served in literal atomic units, an electron at a time, measured in Kilowatt-hour or Joule units.

• Infrastructure: Both AI and electricity need infrastructure to produce valuable output. AI is being served by AI supercomputers in centralized data centers, similar to how power plants are centralized infrastructure for delivering electricity. Both AI and electricity can be generated locally, if you have your own GPU or solar panels, but in both cases, scaling infrastructure is needed to scale outputs.

• Embedded ubiquity: Electricity powers home appliances, vehicles, industrial applications and more. Not a single appliance or application, electricity is the platform for energy delivery for much of modern society. Similarly, AI will be embedded in applications and will be the intelligence in applications and appliances both for personal and enterprise use. AI will be everywhere.

• Economic platform: Electricity was the basis of the second industrial revolution and restructured the economy in the late 19^th and early 20^th centuries. AI is driving yet another technology revolution, accelerating technology-driven growth and disrupting the economy.

These two platforms – intelligence and energy delivered by AI and electricity - are the two main economic pillars of the 21^st century intelligence era economy. All economic value is defined and shaped by the fundamental forces of energy and intelligence.

Everything we make, mine, create, move, fabricate, serve, farm, sell, trade, is based on the intelligence (via decisions, designs, technical specs, artistic inputs, legal determinations) and energy (used in mining, manufacturing, transport, and services) needed to make them happen.

The intelligence pillar is a technology stack built on AI models, training, chips, and data centers; supporting it all is electrical energy itself. It takes many gigawatts to power the supercomputers that run AI workloads. The efficiency of our economy depends on the efficiency of those technology and energy stacks.

AI Progress and Adoption Timescales

The analogy between electricity and AI is useful if the adoption of electricity and its impacts give us clues into how the AI revolution will play out. For example, while there was no electricity ‘bubble’ or bust per se, there were electricity-related booms, bubbles, and busts, as seems to happen with nearly any technology disruption.

The challenge is that AI technology progress and adoption is so much faster than what occurred with electricity. Electricity took 4 decades to achieve penetration above 90% in the US, only reaching it after World War II. Globally, there are still hundreds of millions of people without direct access to electricity.

Since 2023, AI Adoption has grown faster than any major consumer technology on record. In just 2.5 years, generative AI reached a billion global users, which is much faster than smart phones, social media, or the internet let alone electricity.

Verasight reports that roughly 64% of Americans now use AI at least monthly, a curve that appears much steeper than television’s rise in the 1950s and cell phone adoption in the 1990s. Adoption has gone global as well; several other countries such as Singapore report higher AI adoption than the US.

The adoption of generative AI has largely centered on AI chatbots. The billion users have interacted with AI through ChatGPT or other AI chatbot interfaces. The use of AI as agents is even newer, but the curve has been just as steep.

Even faster than AI’s adoption curve has been the rapid advance of AI capabilities. AI models in 2026 are vastly more capable than they were just a few years ago, with many AI labs competing at the frontier and delivering AI models in mid-2026 that all would have been state-of-the-art had they been delivered just 6 months prior.

Rather than slowing down, AI progress is accelerating in 2026 thanks to recursive self-improvement, as AI itself accelerates improvements in the AI technology stack. An example of that has been OpenAI’s GPT-5.6 Sol optimizing their Luna model.

Google’s current troubles show how quickly things are shifting in the AI race. Gemini 3.1 Pro was briefly the best AI model in the world when it previewed last November. Now, Google is a laggard because they haven’t delivered a successor in the past 6 months to their formerly SOTA flagship AI model.

The Five AI Application Types and Five Waves of AI

Just as AI’s killer app was the chatbot, with ChatGPT, electricity found its “killer app” in the lightbulb. The initial application of electricity brought light to homes and workplaces. Once electrical power was available and motors could be adapted to different uses, new inventions arose to plug in to electric power, from the vacuum cleaner to the refrigerator to the factory machine tool. Each electrical invention and application had its own invention, development and adoption path.

As a general-purpose technology like electricity, AI has many potential applications, which can be grouped into five major categories:

• As an AI chatbot to answer queries.

• As an AI creation tool, to generate images, music, video, writing and other content.

• As an AI agent, to perform tasks, generate software, and automatically complete workflows.

• As an AI researcher or AI scientist, to perform scientific analysis, generate mathematical proofs, develop new molecule or drug candidates, or invent new technologies or innovations.

• Embodied as physical AI in an AI robot, self-driving car, or other physical AI application.

These five broad categories of use cases of AI have distinct needs and characteristics while sharing common need for AI intelligence. While we could partition AI applications in more fine-grained detail, the five categories capture the broad scope of what AI can do.

As with electrical appliances, AI’s various applications in agentic AI, physical AI and inventive AI are on different development paths and timelines. The implications of this are that the AI “wave” is not a singular wave, but a succession of waves, each one riding its own adoption S-curve. These are the 5 waves of AI.

Three years ago, in one of my earliest AI articles, I spoke of that time as the “steam engine days of AI,” to express how much more AI efficiency and capability was still to come. We have come far with the development of AI reasoning, massive improvements in AI models, and the huge uptake of AI adoption and agentic AI tools. Yet even now, there is no wall to further AI improvement.

Looking at the landscape of these five AI applications areas today in 2026, AI may have crossed the chasm into general use for some uses, but we are still in early days in terms of AI’s ultimate development, capabilities, adoption, and impact. We are already far along the AI chatbot adoption curve, but that does not mean that AI overall is anywhere near mature; there is less penetration for the other 4 applications, and those other four applications are more impactful and provide more leverage to our economy.

We will explore further the 5 waves of AI and the specific status of each of these AI waves in a follow-up.

Top Tools – Hermes Bot Mode

XAI’s recently released Grok Bot has become very popular for persistent, cloud-hosted multi-agent workflows. Instead of one bot, you create different connected bots to accomplish different tasks in multi-agent way; you can think of them as different AI employees serving different roles for you. Since good AI design patterns get copied, others are replicating how Grok Bot works.

Nous Research released Bot Mode for the Hermes agent. The mode allows Hermes agents with specialized roles to operate persistently and coordinate automated tasks:

Your agent profiles become a series of named Bots. Each Bot has its own role, model, memory, skills and profile picture; Bots can use any model and even communicate with each other.

This adds a great feature to the already great Hermes Agent.

CopilotKit released OpenBot as an open implementation of the persistent multi-agent bot pattern. It follows the pattern of cloud-hosted, always-running agents used by Grok Bot and similar systems and It developers create specialized agents that remain available and coordinate work across tasks.

AI Tech and Product Releases

Ornith released the open-source Ornith-1.5 model family, consisting of a 9B parameter dense model and 35B and 397B parameter MoE (mixture-of-experts) models. These models are based on Qwen 3.5 and Gemma 4 base models and were post-trained through a self-improvement process that generated tasks, constructed agent scaffolds and self-evaluated work. Ornith-1.5 397B scores 86.1 on Terminal-Bench 2.1 and 56.0 on DeepSWE, benchmark results comparable to DeepSeek-V4-Flash-0731. Ornith-1.5-35B-A3B is an efficient AI model for local AI coding agents, on par with Muse Glimmer 30B.

Xiaohongshu’s dots studio released dots3-note preview, an open-weights MoE model with 280B total and 16B active parameters and a 512,000 token context window. It targets long-horizon agentic tasks, and it is multimodal, accepting text, images, video and audio input, with text-only output.

Ant Group’s InclusionAI released the Ling-3.0 MoE model family: Ling-3.0-tiny has 7.9B total parameters with 1.3B active parameters per token; Ling-3.0-flash has 124B total and 5.1B active parameters, while achieving mid-tier results, comparable to MiniMax-2.7. InclusionAI released pretrained, intermediate-training and merged checkpoints, with quantized versions on HuggingFace intended for local deployment.

Cartesia released the Sonic-3.6 text-to-speech model, the streaming speech model that supports 44 languages and can produce audio in under 90 milliseconds latency. Sonic-3.6 reached first place on Artificial Analysis’ Provider Voice and Controlled Voice leaderboards, providing an independent indication of its speech quality.

Alibaba launched HappyShrimp 1.0 for end-to-end music generation, which converts prompts describing a genre, emotion or story into complete songs, generating the lyrics, melody, arrangement, vocals and final audio production. Users can also provide their own lyrics. Independent comparisons of its musical quality and consistency against systems such as Suno or Udio are not yet available.

Liquid AI released new four-bit LFM2.5 checkpoints for edge device models. The 230-million to 2.6B parameter LFM2.5 models were trained with quantization-aware distillation to recover approximately 97% of their BF16 benchmark performance with low memory footprint and higher throughput.

xAI released Grok Build to all users. Grok Build generates working websites, applications, games and dashboards inside Grok conversations on the web, iOS and Android. Users can revise projects conversationally, publish them to dedicated URLs and connect generated applications to xAI APIs for Grok chat, image and voice capabilities.

TrueFoundry open-sourced TrueForge, an open agent harness that supports models from over 20 other providers, along with MCP servers, sandboxed execution, human approvals, persistent sessions, context management and self-hosted deployment. TrueFoundry reported 30% to 75% lower task-completion costs than Claude Managed Agents in its tests, but those results depended on model selection, workload and the company’s own evaluation methodology.

Superwhisper released S1-mini as an open-weights transcript-cleaning model. The 0.6B parameter Qwen3 fine-tune is specialized for English transcript cleanup; it converts raw speech-recognition output into punctuated, formatted written text and can run locally, avoiding sending data to a server. S1-mini has been released in standard and GGUF formats for local use.

Amazon Bedrock expanded access to Grok 4.6 and OpenAI’s GPT-5.6 models. xAI made Grok 4.6 generally available through Bedrock, while AWS added cross-region inference for GPT-5.6 Sol, Terra and Luna across more than 25 regions.

Adobe made its Firefly AI audio tools generally available. Firefly’s AI-based Generate Music, Generate Speech and Generate Sound Effects tools produce background music, voiceovers and precisely timed sound effects inside Adobe’s creative platform. Adobe said the commercially released tools were trained on licensed and public-domain material. The new free Firefly AI Assistant provides limited daily generations.

Salesforce launched Slack Code for Slack-based collaborative agentic programming. Dedicated code channels allowed teams to assign work to agents such as Claude Code and Devin, examine plans and code changes, run previews, provide feedback and approve deployment from a shared Slack workspace. This moves AI coding activity out of private terminals and into auditable team channels, a feature that other agents and applications such as Claude Tag provide.

Anthropic added a /design command to Claude Code. The research-preview command creates and revises Claude Design artboards from Claude Code’s command-line interface or the Claude desktop application. It integrates visual interface design into a coding workflow, allowing developers to iterate layouts before implementing them in code.

Cursor launched Origin, a Git-compatible code-hosting platform that competes with GitHub. Origin integrates repositories, code review, continuous integration and preview deployments with Cursor’s coding agents and was designed to accommodate the high commit volumes produced by multiple agents working in parallel. Cursor also expanded its cloud agents to be able to modify and test repositories in isolated environments and continue working while users were offline.

Chroma introduced Foundation as a persistent-memory service for AI agents. Foundation stores information and outcomes from previous sessions as a self-improving wiki and make them retrievable across tools and AI agents. This avoids having to force each agent session to rediscover the same context and procedures and solves the episodic agent memory challenge.

OpenAI added Apple Messages integration and expanded Computer History in ChatGPT Work. The Apple Messages plugin allows ChatGPT and Codex on Apple Macs to search for iMessage, SMS and RCS conversations and prepare or send messages, with recipient and message approval enabled by default.

OpenAI previewed Private Safety Processing for zero-data-retention customers. In this system, customer content remains on customer-controlled infrastructure or in storage encrypted with customer-controlled keys, and OpenAI receives only narrowly defined safety signals when automated systems detect a risk. The system analyzes patterns across related API interactions to detect misuse or agent behavior that could not be recognized from a single request, while keeping the underlying prompts and responses inaccessible to OpenAI.

Modular released the complete Mojo compiler and toolchain as open source. Mojo is designed as a general-purpose systems language for CPUs, GPUs, AI accelerators and high-performance AI software, combining Python-like syntax with lower-level control. The Apache 2.0 release with LLVM exceptions opened the compiler, tooling and language implementation after Qualcomm completed its acquisition of Modular.

Ramp launched Router, a unified API for accessing AI models. The service allows businesses to switch among multiple AI models from different providers while monitoring token spending, latency and fallback attempts. Router is limited to U.S. customers for now.

Meta expanded its Pocket vibe-coding application to users across the United States. Pocket uses Meta’s Muse Spark model to create small interactive games and other experiences from natural-language prompts and distribute them through a social feed where others could play, share and remix them.

Apple Music is preparing for mandatory labels for AI-generated and materially AI-created tracks. The planned “Made With AI” labels will replace Apple’s earlier voluntary approach and require music distributors to identify qualifying releases.

OpenAI paused frontier AI model Astra development after new cybersecurity warnings arose in preliminary testing of its unreleased Astra model, which suggests it possesses critical cyber capabilities. OpenAI paused reinforcement-learning work on its models for two weeks to focus on strengthening its safeguards. They are adding stronger sandbox and network isolation, trajectory-level monitoring, and continuous security testing.

AI Research News

MIT researchers documented “attribution decay” in diffusion models. Researchers trained 24 diffusion-model ensembles on datasets ranging from 256 to more than 160,000 images and found that removing an individual image, person, or artist increasingly failed to change the generated output as datasets grew. The finding applied to diffusion models and individual-level causal attribution; it did not establish that training datasets were legally irrelevant, eliminate occasional memorization, or demonstrate the same effect in LLMs.

Google DeepMind published an overview tracing 15 years of game-based AI research in games from Atari reinforcement learning to the development of SIMA 2. The foundational techniques developed through game research, including RL, have contributed to development of reasoning AI models and specialist AI models such as Alpha-Fold as well as broader scientific breakthroughs.

An example of AI advancing science is how Claude designed experimentally validated protein binders. Claude autonomously designed 1,320 protein candidates across 15 targets by orchestrating specialist protein-design and structure-prediction models. Laboratory testing found 354 binders against 14 of the 15 targets with interpretable results, representing an overall hit rate of about 27%.

The downstream benefits of AI can be seen in the news that Moderna and Merck reported positive Phase 3 results for an individualized mRNA cancer therapy. AI and automated algorithms reportedly contribute to the sequence design and were part of the design pipeline for this mRNA therapy that involves individualized medicine, showing that AI can accelerate developing new cures.

AI Business and Policy

Stripe agreed to acquire OpenRouter, a gateway that routes AI requests across more than 400 models from over 80 providers, giving developers a unified interface for selecting models, managing availability and controlling token expenditures. Reports valued the transaction at over $8 billion. Stripe said it planned to combine model routing with its token-billing and financial infrastructure while retaining OpenRouter’s multi-model approach. OpenRouter CEO says their mission continues.

Cursor announced that its acquisition by SpaceX had closed and that its team would join SpaceXAI.

Data-labeling startup Micro1 grew its gross annual run rate from $100 million to $500 million over an eight-month period driven by surging demand for AI training data , showing the expanding market for specialized training data as top AI labs scale their models. The company relies on a network of domain experts for model evaluation and reinforcement learning and is expanding into automated synthetic data generation and robotics pre-training datasets.

Ramp released transaction data showing that OpenAI began narrowing Anthropic’s market lead among U.S. business users during the third quarter. Anthropic held a lead over OpenAI in corporate adoption earlier in the year, but OpenAI accelerated its growth following the release of its GPT-5.6 Sol model.

OpenAI cofounder and president Greg Brockman significantly expanded his operational purview to become second-in-command behind CEO Sam Altman. Brockman assumed leadership over the company’s product strategy, super app efforts, and scaling division as OpenAI reorganized ahead of an anticipated initial public offering.

OpenAI joined the PORTS-Pike Technology Data Center project at the former Portsmouth uranium-enrichment site in Ohio. The planned campus would provide as much as eight gigawatts of AI computing capacity under a 20-year lease.

AI Opinions and Articles

Filmmakers Matti Haapoja and Sam Kolder faced public backlash after posting promotional videos for AI video platform Higgsfield and Seedance 2.5. Fans and fellow creators criticized the lack of disclosure of potential paid partnerships and the normalization of generative AI tools in professional filmmaking. The controversy adds to ongoing debates regarding AI-generated content, and it also highlights how AI’s public-acceptance problem has become a business PR problem.

OpenAI announced AI Futures, a new initiative led by Dean Ball focused on long-term policy, economic integration, and institutional governance for advanced AI. In launching OpenAI’s Strategic Futures team blog AI Futures, Dean Ball argues that autonomous systems could allow governments and powerful institutions to operate military, administrative, and economic machinery with less dependence on human labor and public cooperation.

He proposes balancing power among institutions while preserving individual access and autonomy rather than relying on either centralized control or unrestricted decentralization.

TL;DR: This week is chock full of AI releases: Five new frontier-level AI models, several great local AI models, some new audio, image and video AI models, and new harnesses. AI progress is not slowing down but accelerating thanks to recursive self-improvement.

Top Tools: Gemini 3.7 Flash, Grok 4.6, Qwen 3.8 Max,GLM-5.3, DeepSeek V4 Pro 0813

Google released Gemini 3.7 Flash Google’s updated multimodal model supporting text, image, video and speech inputs. Gemini 3.7 Flash improved substantially over its three-week-old predecessor Gemini 3.6 Flash on coding, web development, and business automation, including gains from 49% to 65% on DeepSWE and from 17% to 30.4% on Automation Bench. Artificial Analysis measured it speed at a zippy 340 output tokens per second, and they gave the high-reasoning version an Intelligence Index score of 56, comparable to GPT-5.6 Terra but priced at a much lower $0.75 / $3.75 per million input / output tokens, half of Gemini 3.6 Flash’s original price.

While Google has failed to deliver their flagship Gemini 3.5 Pro, Gemini 3.7 Flash nearly makes up for it with near-Pro intelligence and a compelling price-performance-speed profile for day-to-day agentic AI. Gemini 3.7 Flash is available across Google’s AI products, including AI Studio, API, Enterprise Agent Platform, Antigravity, and Gemini Spark, Google’s personal AI agent.

SpaceXAI released Grok 4.6, their latest frontier model that scores 61 on the Artificial Analysis Intelligence Index, boosting Grok 4.6 as a direct competitor to GPT 5.6 Sol and Claude Opus 5, yet costing far less: $2 / $6 per million input / output tokens. Benefitting from their Cursor team and trained on model-generated reasoning data for long-running agents, interactive coding, and knowledge work, Grok 4.6 scores 69.9% on CursorBench 3.2 and 61.3% on FrontierCode 1.1. Grok 4.6 is now available through GitHub Copilot and the SpaceXAI platforms: API, Cursor, Grok Build, and Grok console.

Chinese AI startup Z.ai updated their GLM-5.2 model with further post-training to release GLM-5.3, a model that now outcompetes Kimi K3 and matches Grok 4.6 for long-horizon tasks. GLM-5.3 features advanced cybersecurity capabilities as well as substantial gains in long-horizon coding. It reportedly discovered a serious vulnerability in Cursor shortly after its launch.

Z.ai’s shared the secret to GLM-5.3’s success: “Scaling post-training is all we did for GLM-5.3.” All recent AI releases by AI labs have followed the same pattern: They are scaling RL post-training by having the AI models themselves generate synthetic training and verification data that trains AI models on long horizon tasks. This recursive self-improvement loop is highly automated and rapidly increasing the rate of AI model improvement for the time being. One sign of the speed of progress: Elon Musk said Grok 4.7 was expected within three to four weeks of the Grok 4.6 release.

Alibaba published the open weights for Qwen 3.8 Max as Qwen3.8-2.4T-A95B on HuggingFace, making this Qwen’s largest open-weight model to date. It features 2.4T total parameters and 95B active parameters in a hybrid full-and-linear attention architecture, allowing a native 262,144-token context that can be extended to approximately one million tokens. The model includes configurable reasoning controls and a fine-grained mixture of experts design to balance performance and serving costs.

Nvidia optimized the model to run on its GB300 NVL72 platform, delivering high throughput for demanding reasoning and agentic workloads.

DeepSeek officially released V4 Pro 0813, their updated flagship agentic AI MoE model with 1.7T total parameters, 49B active parameters and a one-million-token context window. The model scored 87.9% on Terminal-Bench 2.1 and 62.7R on DeepSWE, substantially improving on the V4 Pro preview. Its weights are available for local deployment under the MIT license and API via third parties including OpenRouter, while DeepSeek on its API platform replaced its flat API pricing with higher peak and off-peak rates.

AI Tech and Product Releases

Qwen team released Qwen3.8-27B, a game-changing open-weights local AI model yet that is SOTA for its size, achieving performance comparable to Opus4.6 Max. For example, Qwen3.8-27B achieves 42% on DeepSWE 1.1, 73% on Terminal Bench 2.1, and 84.3% on OSWorld-Verified. It supports native video and image understanding and adjustable reasoning effort, This must-have local AI model is available via HuggingFace that can be used for agent execution in popular harnesses and development tools.

Meta introduced Muse Glimmer, a 30B parameter dense open-weights model released under the Apache 2.0 license and optimized for multimodal understanding, tool use and coding for local on-device AI. Benchmarks show Muse Glimmer performs better than Gemma4 31B, comparable to Qwen3.6 27B, but not as intelligent as just-released Qwen3.8 27B. Muse Glimmer can run on consumer devices using weights available via Hugging Face, needing only 14 GB to 18 GB of GPU or unified memory when using dynamic 4-bit quantization.

ByteDance officially released the Seed2.1 model family, Seed2.1 Pro and the faster Turbo variant targeting complex productivity and software-engineering. ByteDance reports benchmark results for Seed2.1 Pro showing comparable results to Gemini 3.1 Pro and GPT-5.5, but the evaluations are mostly company-selected, and the release did not provide open weights or third party evaluation.

Nvidia released Nemotron 3.5 Lightning, an open-weights 30B parameter MoE model with 3B active parameters designed with the Mamba-2 Attention hybrid architecture. It supports context windows of up to one million tokens and is optimized for long-running autonomous agents, subordinate-agent workloads and local inference. Like Qwen3.8-27B and Muse Glimmer, Nemotron 3.5 Lightning has weights available on Hugging Face and supports local deployment through Ollama, llama.cpp, vLLM and TensorRT-LLM.

Nvidia’s Lightning 3.5 model release included a number of speculative decoding methods for faster text generation: DSpark, DFlash and multi-token-prediction decoding. Additionally, Nvidia launched NeMo Switchyard, an open-source AI routing library that dynamically allocates workflows between frontier and efficient models in real time. Together, these tools cut benchmark execution costs to a fraction, significantly lowering the expenses of running enterprise AI agents.

OpenAI previewed GPT-5.6 Sol Ultrafast, a service tier that runs on Cerebras hardware at 750 output tokens per second, up to 14 times the speed of standard processing. The service is intended for latency-sensitive work such as incident response, financial market analysis, voice applications, customer support and live research. The feature is currently available to a select group of customers and will expand as infrastructure capacity grows.

OpenAI introduced GPT-5.6-Cyber, a version of GPT-5.6 Sol trained for authorized vulnerability research, exploit validation, zero-day discovery and advanced security testing. The model is available only to approved cyber-defender security partners like Cisco and Cloudflare through the Daybreak Red access tier. GPT-5.6-Cyber achieves a 95% completion rate on its advanced cybersecurity prompt evaluation, compared with 1.5% for standard GPT-5.6 Sol. During testing, the model identified critical zero-day vulnerabilities, including memory-corruption exploits in the V8 JavaScript engine.

SpaceXAI released a persistent AI agent called Grok Bot designed to perform various knowledge work tasks. Grok Bot is available on desktop (Windows and macOS) and iOS and runs tasks and AI models in cloud sandbox environments. Users can assign multiple bots to parallel tasks, connect them to applications and websites, and teach them repeatable routines while retaining approval for sensitive actions. Grok Bot is in early beta release and is included with Cursor Ultra and SuperGrok Heavy.

DeepSeek also released DeepSeek Harness, a modular open-source AI agent harness designed to execute coding tasks and long-running multi-step agent workflows. DeepSeek Harness is designed to compete with developer tools like Anthropic’s Claude Code. The harness includes a web interface and allows developers to add models, tools, skills and other components as plugins to the harness, presenting nearly every part of the agent runtime as a plugin. The repository accumulated approximately 23,000 GitHub stars within days of release.

South Korea’s Motif Technologies released Motif 3 under the MIT license with 314 billion total parameters and approximately 13.2 billion active parameters. The mixture-of-experts model is designed for agentic work, tool use, coding, reasoning and general knowledge tasks. Its developers reported a 76.2 percent result on SWE-Bench Verified and also released a corresponding base model.

Cohere released North Micro Vision Instruct, a 2.4B parameter open weight vision-language model under the Apache 2.0 license. It supports native-resolution images, multiple images and multilingual prompts for visual question answering, captioning, grounding, OCR, chart analysis and document processing.

Liquid AI released LFM2.5-VL-3B, an open weight 3B parameter vision-language model designed for on-device image, video, OCR and visual-grounding tasks. The model uses about three gigabytes of memory and was measured at 228 tokens per second on an Apple M5 Max and 116 tokens per second on an AMD Ryzen AI Max+ 395. Liquid AI also reported that it can run fully on a Galaxy S26 Ultra at approximately 20 tokens per second.

SpaceXAI released DeepSeek Harness for image generation and editing through Grok’s Quality Mode and its API. The model is designed to follow detailed instructions, render typography and complex layouts, and preserve subjects and other elements across repeated generations and edits. Grok Imagine Image 2.0 ranks second in Arena evaluations for text-to-image generation and image editing.

Lightricks released LTX-2.5, an open-weights 22B parameter diffusion model for text-to-video and image-to-video generation that offers fast 1080p high-quality audio-video generation. The updated LTX-2.5 model includes a new diffusion video decoder and a custom Gemma 4 language backbone, and it supports multi-shot sequences, synchronized audio and video, first-frame and last-frame controls, and fine-tuning with user data. LTX-2.5 integrates natively into ComfyUI and runs locally on Nvidia RTX GPUs with 16GB minimum memory, with fast real-time rendering (10 seconds of 1080p AI video in 6.8 seconds) for cinematic and physical AI applications. Model weights are available for download, or it can be run via a managed API tier.

Alibaba released Wan-Animate-2, a 14B parameter diffusion-transformer model for character animation under the Apache 2.0 license. The model accepts a reference image, a driving video and a text description to transfer motion while preserving the character’s appearance. Alibaba also released a distilled configuration that performs generation in ten inference steps without classifier-free guidance.

MiniMax released Music 3 as an open-weight model for producing complete songs from lyrics and detailed musical descriptions. It can generate tracks of up to five minutes with vocals, evolving arrangements and sustained musical structure. The model is intended for production-oriented music creation and is available through Hugging Face.

Google DeepMind introduced a massively multilingual sign-language-to-text translation model called SL2T. The breakthrough model powers new sign language features in Live Transcribe on Pixel devices, initially supporting American Sign Language translation. The system overcomes core translation and computer vision challenges to enable deaf and hard of hearing users to interact with devices using natural sign language.

Microsoft released MAI-Code-1.1-Flash, a lightweight, agentic coding model integrated into GitHub Copilot and VS Code. The new model delivered higher code quality with 25% greater token efficiency and at one-quarter of the cost of its predecessor. It achieved notable performance gains on terminal and .NET development tasks while speeding up token streaming for developers.

Anthropic updated the biology safeguards for Claude Fable 5, reducing false-positive fallback rates by approximately 85% across product surfaces. The adjustment allowed the AI model to assist users with a broader range of everyday health, educational, and clinical biology tasks without unnecessarily switching to a less capable model. Some high-risk dual-use queries involving virology, toxicology, and molecular design continued to be restricted pending trusted access pathways.

Writer introduced Palmyra X6 and a more efficient agent harness. Built through post-training GLM-5.2 model, Palmyra X6 was optimized for marketing, sales, and other enterprise workflows. By combining this model with upgrades to its agentic harness infrastructure, Writer reduced operating costs and improved speed by over 40%. Additionally, the release introduces new governance tools to help organizations efficiently manage per-task expenses in enterprise workflows.

Anthropic rolled out a capability update to Claude Tag for Slack, implementing full-channel context memory, predictive participation heuristics, and proactive collaboration modes across its Enterprise plans. The updated architecture improves the model’s judgment regarding when to intervene in technical discussions without explicit mentions.

Anthropic began applying invisible watermarks to text from new Claude models and said it was working to extend support to earlier models. Anthropic introduced the measures to meet EU AI Act transparency requirements and said it planned to provide detection documentation and tools. The policy applies worldwide across Claude products. Requiring watermarks in users’ output, even if it is not visible, has led to consumer backlash and subscription cancellations.

Google now lets users toggle off visible watermarks on AI-generated images, videos, and music from models like Nano Banana, Omni, and Lyria. While the visible stamps are removed, content continues to use invisible SynthID watermarks and C2PA metadata for background verification.

AI Research, Business and Policy News

Due to many AI releases this week, we don’t have space for other news.

Top Tools – Qwen 3.8-Max and Muse Spark 1.2

Alibaba officially released Qwen3.8-Max, the 2.4T parameter flagship Qwen AI model with 95 billion active parameters positioned for coding, coworking, and extended agentic tasks. Qwen 3.8-Max accepts multimodal inputs and supports a 1M token context window. In a sign that Chinese AI labs are returning to the open release AI strategy, Qwen promises to release weights for this AI model soon.

Qwen 3.8-Max achieves frontier-level benchmark scores on coding (67% on SWE-Pro) and agentic tasks (86% on OSWorld-Verified) and an Artificial Analysis overall intelligence score of 56; that score places it the world’s fifth best AI model, just behind Kimi K3. Early reviews tend to confirm its high performance. API pricing is a highly competitive $2/$6 per million input/output tokens.

Meta released Muse Spark 1.2, a coding-focused update to Muse Spark 1.1, alongside Muse Code, a terminal coding agent designed for repository-scale software engineering. Coming just 27 days after its predecessor, Muse Spark 1.2 features a 1 million token context window and boasts significantly improved intelligence, especially on coding and agentic tasks, scoring 82.9% on Terminal-Bench 2.1, 59.3% on Deep-SWE. Its overall Artificial Analysis intelligence score is 54, just below Qwen 3.8 Max and on par with Grok 4.5.

Muse Spark 1.2 was co-trained with Muse Code to optimize performance for long-horizon coding tasks, including whole-repository generation, debugging, and iterative GPU kernel optimization. Unfortunately, the AI model is proprietary, accessible only through the Meta Model API, but the API pricing is a low $1.25 / $4.25 per million input / output tokens.

Meta beta release of their Muse Code terminal-based coding agent brings features designed to plan, write, and validate code changes across large repositories. Powered by Meta’s Muse Spark coding model, the system can divide large assignments among multiple sub-agents working simultaneously in isolated work trees, preventing them from directly modifying or conflicting with the user’s active working copy.

Meta and Qwen are in the frontier AI race, and there are rumors of Gemini 4.0 and impending releases of GLM-5.3, and Grok 4.6 coming soon. The bigger picture is that AI competition is fierce and there is no wall to improving performance.

AI Tech and Product Releases

Minimax launched Minimax H3, an open weights omni-modal model for general purpose audio-video generation. Minimax states:

H3 can jointly understand multimodal contexts spanning text, images, video, and audio. It generates video with native stereo audio at up to 2K resolution and 15 seconds in length.

Minimax H3 is a strong model, achieving number two spot on Arena for image-to-video generation. Pricing for Minimax H3 is $0.13 per second for 2K generation, and $0.09 per second for 768p.

Nvidia released Alpamayo 2 Super, an open 34B parameter reasoning vision-language-action (VLA) model designed to accelerate autonomous vehicle development. The model combines the 32B parameter Cosmos 3 Super Reasoner with a 2B parameter diffusion-based Action Expert and is post-trained with reinforcement learning to handle rare driving scenarios with inspectable decisions. This multi-task architecture provides autonomous vehicle developers with a common foundation for trajectory generation, reasoning traces, and data labeling.

Google expanded Ask Maps with tools that can identify restaurants and assemble food orders through services such as Toast, Square, and Uber Eats. The agentic AI assistant can also compare hotel prices and availability, find event tickets, remember previous conversations, and can draw from users Personal Intelligence for personalized decisions. The features are initially rolling out in the United States.

MacPaw partnered with Liquid AI to develop Elix, an on-device inference system for locally hosted AI models, along with a local memory system for MacPaw’s Eney assistant. MacPaw plans to make the infrastructure available to Setapp developers, enabling applications and agent workflows that can operate offline while keeping more user processing on the device.

AWS announced the general availability of Web Search on Amazon Bedrock, a server-side built-in tool that grounds foundation model responses in current web knowledge. The feature combines an continually refreshed web index with a built-in knowledge graph to deliver semantic snippet extraction and reduce factual inaccuracies.

AWS also released an automated web insight extraction solution utilizing Amazon Bedrock AgentCore Browser, Amazon Bedrock, and Amazon OpenSearch Serverless. The event-driven architecture uses a managed browser service to generate AI-powered summaries from JavaScript-heavy web pages and RSS feeds reliably.

Google made Gemini in Classroom available to K-12 and higher education students of all ages. The integration enables students to generate interactive study tools from course assignments using contextualized starter prompts and to sync educational materials to Gemini Notebook.

Asana unveiled Agentic Work Management (AWM), an operating system built on top of its Work Graph architecture that treats AI agents as collaborative teammates. The system enables agents to share company-wide memory and automate complex workflows while enforcing strict access controls to prevent data leaks.

Wrinkles is a newly released iOS and Android application that functions as an AI-powered audio tour guide based on user location. The app draws from a global database of 1.3 million historical points of interest across 177 countries, allowing users to discover local history and add personalized memories. Users can also interact with the platform through conversational questions and share custom maps with friends and family.

Google announced that it will begin removing access to Google Assistant on Android phones, tablets, and paired devices starting September 4th. Users in regions where Gemini is available will transition exclusively to the AI-enabled Gemini across mobile devices, smartwatches, headphones, and Android Auto.

AI Research News

OpenAI solved ten major open mathematics problems using an internal version of its forthcoming Astra AI model. These results represent a transition to AI systems that can contribute genuinely new mathematics: Generative Math. We discuss this accomplishment and its implications in depth in our prior article “OpenAI’s Astra Tackles Mathematical Invention.”

SaferAI published a report revealing that Z.ai’s open-weight GLM-5.2 model exhibited no refusals for offensive cybersecurity or biological tasks during evaluations. The Safer AI report found that GLM-5.2’s cyber capabilities were roughly two to four months behind leading closed models and that its biological performance was comparable to GPT-5.5 and Claude Opus 4.7, but GLM-5.2 refused none of the offensive-security or biological tasks presented during testing. The findings demonstrate a widening safety gap as open-weight models approach the capabilities of closed frontier systems without equivalent API safeguards.

This does not mean proprietary AI models are safe. During 122 cybersecurity-evaluation runs involving seven models, the UK AI Security Institute recorded ten runs in which agents took unauthorized actions on the live internet, producing 19 documented actions. Seventeen involved Anthropic’s Mythos 5 and two involved an OpenAI GPT-5.6 Sol configuration with cyber safeguards disabled. One agent created false identities and attempted to persuade an open-source maintainer to approve malicious code, but the attempt failed, and the investigation found no resulting real-world harm.

AI Business and Policy

Alphabet restructured Google DeepMind leadership. Demis Hassabis moved from Google DeepMind CEO to chairman of the unit and Alphabet chief scientist, while DeepMind CTO Koray Kavukcuoglu assumed day-to-day operational responsibility.

Meanwhile, Jeff Dean, Sanjay Ghemawat, Oriol Vinyals, and Quoc Le left Google to establish Discovery Loop, a public-benefit company focused on automating machine-learning, scientific, and engineering experiments. Alphabet invested in the startup and Google Cloud agreed to supply computing capacity.

Agentic AI use is skyrocketing. Shopify said traffic and orders reaching its merchants through AI services tripled year over year during the second quarter. They described AI discovery as complementary to conventional search rather than a replacement, with smaller merchants benefiting when AI services direct shoppers toward products from a wider range of online stores.

Anthropic agreed to buy six years of cloud capacity for $10 billion from AI infrastructure startup Volta. The planned 133-Megawatt Norway data center would be developed with Bitdeer and powered by Nvidia Vera Rubin systems.

AI research company Mirendil signed a multiyear Google Cloud agreement valued at more than $100 million. The arrangement provides access to Google TPUs, Nvidia GPUs, and managed training clusters for Mirendil’s work on systems intended to conduct continuing research and improve their knowledge and performance over repeated iterations.

AI infrastructure company Runware launched the Sonic Inference Pod, a modular and transportable data center designed to deliver high-performance AI inference at a lower cost. The closed-loop cooling units require no water and can be deployed rapidly across global sites to offer distributed compute capacity closer to end users. Runware already deployed ten pods across the United States, Europe, and Asia-Pacific to support clients such as Higgsfield AI and Wix.

The large and growing AI data center buildout is raising financial concerns. Kansas City Federal Reserve Bank President Jeff Schmid said the scale and financing structure of AI infrastructure development should be monitored for possible systemic financial effects. He called for macroeconomic discussion about whether the industry’s concentration and capital requirements could eventually create institutions considered too important to fail.

According to Axios, the White House has developed a confidential, voluntary process for reviewing advanced closed models with significant cybersecurity or national-security capabilities before release. The reported framework excludes open models, calls for approximately 30 days of prerelease government review, and would require models to be stored in secured environments with detailed access logs.

The European Commission and national regulators began enforcing a broader portion of the EU AI Act on August 2. New transparency provisions require interactive systems to tell users when they are communicating with AI, require deep fakes to be labeled, and call for AI-generated or substantially altered material to carry machine-readable markings.

Horizon3 raised a $250 million Series E at a reported $2 billion valuation, more than tripling its valuation in fourteen months. Its NodeZero platform uses AI to conduct authorized, continuous penetration testing across live enterprise networks, identifying exploitable pathways without taking production systems offline.

The Open Secure AI Alliance launched a working group called the Shared AI Findings Exchange to establish cybersecurity guidelines for artificial intelligence systems. Operating under the Linux Foundation, the group brings together over 120 technology companies to share incident reports and secure agent infrastructure. Participating organizations have also contributed various open-source vulnerability scanners and agent-building tools to bolster ecosystem defense.

Spotify announced during its second-quarter earnings call that music licensing partner Merlin has joined Universal Music Group in supporting its upcoming AI music remix and cover product. The platform will enable fans to generate covers and remixes using consenting artists’ music while providing formal credit and financial compensation. Spotify plans to launch the feature as a paid add-on for a subset of users in an initial research preview.

Texas Governor Greg Abbott mandated that all new data center projects undergo mandatory audits by the Public Utility Commission of Texas and the Electric Reliability Council of Texas. The state’s grid operator is currently tracking 474 gigawatts of proposed connection requests, with data centers accounting for approximately 90% of the total queue. The regulatory action aims to address potential grid strain, rising electricity prices, and resource consumption driven by infrastructure expansion.

Apple sought a preliminary injunction in its trade secrets lawsuit against OpenAI to prevent the AI developer from utilizing stolen technology in upcoming hardware products. The iPhone maker filed new court documents claiming that its ongoing investigation revealed up to 11 additional former employees may have been involved in mishandling confidential data. OpenAI publicly dismissed the allegations as baseless, asserting that it neither possesses nor desires Apple’s proprietary information.

AI Opinions and Articles

Who is responsible for when an AI agent goes rogue? U.S. computer-crime law offers little precedent for assigning responsibility when an AI agent accesses a system without authorization but lacks the human intent normally required by criminal statutes. Legal experts suggest that companies operating the models could nevertheless face negligence claims if they failed to isolate evaluation environments, define authorized targets, monitor model actions, or maintain reasonable safeguards.

The incidents and legal precedents suggest that agent developers may ultimately be held responsible through conventional product-liability, cybersecurity, or negligence doctrines rather than laws written specifically for AI.

OpenAI Cracks Ten Major Open Math Problems

One clear sign of AI progress is the shift in what constitutes news. A year ago, a Gold medal at the Math Olympiad was news. Now, math benchmarks are saturated, and this weekend, OpenAI announced that their next-generation frontier model Astra solved ten major open math problems.

OpenAI reported that an internal version of its forthcoming Astra model generated solutions or substantial progress on ten long-standing problems in mathematics. They showed their work; each proof solution was subsequently prepared by humans and formally verified with Lean certificates:

We’re releasing 10 such Astra proofs, complete with lean certificates and CoT walkthroughs for each of them.

The 10 results are published in a lengthy 249-page paper, and cover a range of topics in mathematics, across sphere packing, coding theory, group theory, circuit complexity, quantum games, lattice cryptography and extremal combinatorics:

• Improving a 1978-era lower bound on high-dimensional sphere packing.

• An explicit construction of a non-sofic group, open since Gromov’s 1999 introduction of soficity.

• A disproof of Connes’s rigidity conjecture that certain groups are uniquely determined by their von Neumann algebras.

• Resolutions of several Erdős problems: #146, #180 in extremal graph theory; and #183, proving a super-exponential lower bound for multicolor triangle Ramsey numbers.

• Resolving Ehrhart’s volume conjecture: Determining, in every dimension, a bound on the maximum possible volume of a convex body whose centroid is its only interior lattice point.

• A result on quantum parallel repetition, relevant to post-quantum cryptography.

• A new lower bound on algorithmic circuit complexity, and closest-vector-problem hardness.

Don’t worry if you don’t understand the problems or the jargon; most of us non-mathematicians don’t either. Nor can we attempt to understand or verify the paper or proofs; the professional mathematicians are checking this work using LEAN. An example is the proof that The Erdős–Simonovits degeneracy conjecture fails at every level.

OpenAI got these results with Astra, their next-generation AI model that may become GPT-6 and/or a Mythos / Fable 5 equivalent AI model. They claimed the cost was only $2000 in Sol equivalent tokens, a low compute cost for 10 meaningful novel math results.

However, this doesn’t include the human-level preparation plus independent verifiability via Lean, and OpenAI’s report left open a few questions: How much human guidance was there? Who decided which problems to pursue? Did OpenAI cut the AI loose on many more problems, but only reported successful proofs? Was it 1,000 problems and got 1% success rate with 99% failures, or was it more targeted?

Unanswered questions and hype aside, it’s still a huge accomplishment for OpenAI and AI generally. It’s been noted that several proofs are counterexamples to long-held beliefs, which increases their value as it changes understanding in these domains.

Generative Math

Astra’s ten math proofs will still require detailed scrutiny and validation by independent mathematicians, but it shows next-generation frontier AI models are moving beyond solving benchmark-style mathematics toward producing original, formally checkable research contributions.

What was achieved here is what we can call Generative Math. In the same way AI image generation models create novel images, AI coding models craft new software, and AI music generation generates new songs, the Astra AI model generated new mathematical proofs, all of them complex, difficult, novel, and useful.

Generative Math is not just ‘doing harder math’ as a brilliant PhD-level math student at a higher-level IMO, it’s AI engaged in mathematical invention. This marks a new milestone for AI mathematic reasoning.

Invention is implicit in generative AI models. The very randomness that plagued early LLMs and led to problems such as hallucinations makes them just random enough to be inventive as they generate.

However, randomness alone does nothing but stumble in the dark. It takes a keen amount of reasoning and domain-specific intuition to avoid wasting time and to productively explore a huge problem space. Math prodigies and chess grandmasters possess that skill of distilling an enormous search space into the essential mental model and reason with high focus on what matters.

OpenAI put this new level of capability on their 5 stages of AI roadmap some time ago, putting “Innovators, AI that can aid in invention” on the stage beyond Agentic AI. We observed that the latest level of AI capability is long-horizon multi-agent AI, which is consistent with the long-horizon reasoning needed for AI invention. The two capabilities coincide.

Getting to Math AGI

Even if OpenAI’s report doesn’t report on the human effort and scaffolding to make this happen, the plain fact is that these problems have remained open for decades because they are hard. As Pavel on X put it:

I’ve spent well over 10,000 hours studying math in my life, yet I can’t understand these proofs, at least not without weeks of digging deep into each topic. What’s more, none of my math PhD friends know much about these problems either, and they can’t verify most of them without working directly in the field (yes, math is VERY diverse). LLMs are getting smarter than the experts themselves, and I’m not sure we have enough bright human minds to verify everything that will come out of them in the coming years. Remember when we compared AI intelligence to PhD students? I think we’re past that.

If we are past asking if AI is as good as PhD student, how good IS this work? Thought experiment: Imagine if these results were the work record of a young math researcher. Ten profound results that solve long-standing problems in mathematics in a matter of months. Such a prodigy would be called an Einstein-level genius and be on track to win a Fields medal. That sounds like math superintelligence.

Some mathematicians have stated that the results are genuinely significant and more advanced than prior AI math results, but they are not revolutionary in kind. AI can excel at induction/deduction within existing conceptual frameworks, but cannot yet perform the “abductive jump” of inventing new conceptual worlds:

Current LLMs can fill the gaps of knowledge left by humans. But they can’t jump beyond the external boundary of existing knowledge.

This is likely true. The AI has been trained on existing patterns of reasoning. It doesn’t invent new forms of reasoning; it’s real superpower is hammering with patterns of deep reasoning on a problem at a level that goes long past human endurance and patience. There’s a lot of space in the gaps of human knowledge where such dogged effort can yield results.

We might back off and call this gap-filling Generative Math to be closing in on Math AGI, accomplishing useful human-level math research with AI. We can distinguish between types of human-level AGI this way:

• Math AGI: Capable of the generalized work of a professional mathematician.

• Software AGI: Capable of the generalized work of a professional software engineer.

• Virtual knowledge-work AGI: Capable of the work of a knowledge worker working in an online environment.

• Robotic AGI: Capable of the embodied activities of a human being or worker such as a nurse, bricklayer, chef, factory worker, or manual laborer.

We are near Software AGI and Math AGI because the reward signal for RL training in the math and coding domains is clearest, and coding has been a lucrative area for AI labs to focus on.

An AI skeptic might admit AI shines in math because it has a measurable and verifiable reward attached to each problem, while claiming that such verifiability defines AI’s limits. Gary Marcus is correct that “Expertise in one domain does not at all guarantee expertise in all or even most domains.” Progress in math does not automatically translate to progress in other areas of science that are less amenable to clear verifiable rewards.

However, progress in generative math does translate into progress on rigorous deep reasoning, a useful skill of pruning extremely complex solution spaces to productively reason on it. If AI models crack that “rigorous deep reasoning” skill, it could be a key to superintelligence and will broaden its applicability across domains.

The Math Singularity and Industrialized Math

AI intelligence unlocks AI automation when it has a supporting AI ecosystem of memory, skills, agentic harnesses, tool use, and more. Generative Math with AI unlocks a new kind of mathematics development when it has the supporting infrastructure as well.

The leading mathematician Terence Tao gave a recent talk on what this might look like. His views, prior to OpenAI’s announcement, was that frontier AI had gotten to a “junior human co-author” level of capability in math. As stated in this summary of Tao’s views on AI and mathematics:

Tao doubts that genuine “artificial general intelligence” is within reach of current tools but argues a weaker and still very valuable “artificial general cleverness” is becoming real: the ability to solve broad classes of problems by ad hoc, often brute-force or stochastic means that are fallible and uninterpretable, yet succeed at non-trivial rates when coupled with strong verification. … today’s tools are best viewed as stochastic generators of sometimes-clever, often-useful outputs — which yields the characteristic “useful yet unsatisfying” feeling of a magic trick once explained.

Tao paints a scenario of AI as an efficient complement to mathematicians in a ‘factory-style’ big mathematics. Just as Supercolliders use power and brute force to understand particle physics, mathematicians can harness AI to brute-force solve many aspects of mathematical understanding. The AI magnifies the mathematicians’ productivity but changes their efforts to manage and curate how the AI’s focus, akin to deciding what Supercollider experiments to run.

OpenAI has shown that Astra has a deep domain understanding of mathematics that translated into significant novel findings. DeepMind has produced similar profound results in other areas of science. There will be more to come. AI will disrupt and accelerate many processes of research, discovery and invention in mathematics and science.

However, it will be neither instant nor complete. This is not a ‘singularity’. Instead, mathematics and science research is becoming industrialized; AI is making science research cheaper and more scalable at a level akin to what early 1800s industrialization did to textiles. The acceleration this implies is enormous.

AI will play ever larger role in researching, proving and verifying mathematical hypotheses, but like with industrial factories, man and machine will both play a role, leveraging capabilities with AI. Output could accelerate by an order-of-magnitude. In coming years, hand-crafted AI-free science and math may become an artisan activity.

Astra – To the Stars

It is not the most significant advance in mathematical history, but this milestone is both important and shows a clear direction. AI for mathematics has gone from solving 8^th grade level math in 2023, acing high school level math benchmarks in 2024, achieving Gold-level in IMO math in 2025, and now solving major novel mathematical open questions in 2026. We’ve come very far, very fast.

OpenAI has shown with their 10 AI-generated proofs that AI for Generative Math is here to stay. Astra’s concrete, checkable advance in AI reasoning indicates that AGI-level AI for Math is on the horizon. This advance in AI will automate significant parts of mathematical research and accelerate mathematical progress tremendously.

OpenAI has shown that their new Astra AI model is capable of levels of deep reasoning that could unlock not just mathematics, but other areas of science and engineering invention via deep reasoning. Whether and how this will spill over into other domains is unclear, but major AI labs have other areas of science on their radar.

AI will eventually take them on, and the question won’t be if AI accelerates scientific progress, but how much and how soon will it do so.

Top Tools

ByteDance’s latest video model SeeDance 2.5 has officially launched, upgrading its joint audio-video model to generate 30-second clips in a single pass for longer, more controllable AI video. The model can use up to 30 images, 10 videos and 10 audio clips as references. It has precise editing controls with timestamp-specific prompting and editing, camera-perspective controls, green-screen replacement and improved motion, lighting and audiovisual quality. SeeDance 2.5 can also extend videos across multiple rounds while preserving characters, environments and narrative pacing.

User reviews suggest imperfections in complex action sequences, and ByteDance acknowledges that multi-subject interactions can still be unstable. Another limitation is 720p output, less than some video generation alternatives. Curious Refuge calls it “incredible, until it isn’t.” However, Seedance 2.5 is moving AI-generated video beyond isolated clips toward an end-to-end system for creating and revising coherent, multi-minute audiovisual stories.

AI Tech and Product Releases

DeepSeek released DeepSeek V4 Flash Official API in public beta, and they are presenting incredible near-frontier benchmarks for this 284B parameter MoE model with only13B active parameters. DeepSeek V4 Flash 0731 benchmark scores are surpassing V4 Pro-Preview’s scores, including near-SOTA 82.7% on Terminal Bench 2.1 and 54.4% on DeepSWE. In addition, DeepSeek reports:

We’ve massively upgraded its Agent capabilities …The official V4-Flash now natively supports the Responses API format and is fully adapted for Codex!

A limitation of DeepSeek V4-Flash 0731 is that it is text-only, but it is priced at just $0.14 / $0.28 per million input/output tokens, making it the best current cost-performance AI model for many coding and agentic uses, as much as one tenth the cost of comparable closed AI models.

OpenAI announced updates to GPT-5.6 model family, significantly reducing prices for the Luna and Terra models and improving the GPT-5.6 Sol Fast mode to deliver up to 2.5 times faster speeds for Sol without sacrificing intelligence. The most significant improvement is the 80% drop for GPT-5.6 Luna, now only $0.20 / $1.20 per million input / output tokens, making Luna a price-performance leader with DeepSeek V4 Flash.

Open AI also dropped GPT-5.6 Terra prices to $2/$12, a 20% drop, and GPT-5.6 Sol Fast mode in the API now offers up to 2.5x the speed for 2x the price at the same intelligence. OpenAI claims these all came from the “recursive self-improvement” of GPT-5.6 Sol examining their inference infrastructure and implementing efficiency improvements that reduced serving costs. As Chubby said, this means “the speed of releases is increasing, and models are improving even faster.” AI improvements are not slowing down.

Google DeepMind has launched Gemini Robotics 2, releasing three versatile foundational models to support embodied intelligence and robotics. Gemini Robotics ER 2 is a “high-level brain for robots,” an embodied reasoning model for robotic systems that understands the physical world and orchestrates multi-step tasks by planning quickly and handing off execution to a VLA model. The model can also collaborate by processing live video streams and interact with external tools like Google Search.

Gemini Robotics 2 is their latest vision-language-action model (VLA) that converts vision and language input into motor control. It translates natural language into coordinated whole-body movements across diverse hardware like humanoid and robotic arms. The third embodied model is Gemini Robotics On-Device 2, built to run locally on robotic devices that can adapt to run on a broad range of robots.

Demonstrations show a humanoid robot using these models to manipulate objects, perform dexterous tasks such as tying bags, and collaborate with another robot. It’s impressive, but not as impressive as Uniubi’s back-flipping robot dog and not yet production-ready. Google acknowledges that movement speed and multi-finger reliability still need improvement. Gemini Robotics 2 models are publicly accessible to developers via the Gemini API, Google AI Studio, and the Gemini Enterprise Agent Platform.

SpaceXAI announced Grok Voice Think Fast 2.0, their next-generation intelligent voice model with improved conversational capabilities and state-of-the-art transcription accuracy, especially in noisy real-world environments. This model can reason while speaking, making the model substantially smarter than other speech-to-speech models with no impact on latency. It is priced at $0.08 per minute of audio and integrates reasoning tokens to execute tool calls more efficiently during live conversations.

Microsoft launched MAI-Cyber-1-Flash inside MDASH, putting a model built to find challenging software vulnerabilities into Microsoft’s MDASH vulnerability identification and remediation harness. Microsoft’s first dedicated cybersecurity model is designed to handle roughly 90% of vulnerability-analysis tasks efficiently, while MDASH routes the hardest cases to larger models such as GPT-5.4. Microsoft reports that this multi-model system scored approximately 96% on CyberGym, 12% better than even Claude Mythos while costing 50% less than its previous MDASH configuration.

Last week, Block launched Buzz, an open-source AI Agent collaboration platform that combines Slack-like channels, threads, direct messages and voice with code repositories, workflows and first-class AI-agent participation. Since then, Buzz has hit 20k stars on GitHub and generated buzz as a truly AI-native collaboration layer, not just another agent harness.

Built on the decentralized Nostr protocol, Buzz gives every human and agent a portable cryptographic identity, permissions and auditable activity record, while its ACP integration connects existing agent harnesses, including Codex, Claude Code and Goose, to the shared workspace. This could be a preview of a new AI orchestration paradigm, but it remains an early version, with mobile clients and huddle features still under development. Consider it for hobbyist exploration; it’s not a “Slack killer” yet.

Thinking Machines released Inkling-Small, an efficient open-weights 276B parameter Mixture-of-Experts model with 12B active parameters that achieves comparable performance to their Inkling model at a quarter of its size. Inkling-Small is a natively multimodal model crafted for audio intelligence, making it suitable for real-world audio applications. As with Inkling, Inkling-Small is designed for enterprise developers who customize the AI model for specific uses.

SpaceXAI introduced an integration that brings its Grok AI assistant directly into Google Workspace applications, including Sheets, Slides, and Docs. The tool enables users to query spreadsheets, generate presentations from outlines, and draft text documents using natural language commands. The add-on is available for free on the Google Workspace Marketplace alongside its existing Microsoft 365 integration.

Google updated Gemini Spark with direct Google Chrome integration for direct web browsing use, allowing the AI assistant to use logged-in accounts and saved passwords to handle complex web tasks like scheduling appointments and researching flights The updated features include built-in safeguards against prompt injection and require user approval for sensitive actions such as payments. Additionally, Google expanded access to Gemini Spark for Google AI Pro subscribers, putting the Spark agentic interface inside the Gemini web and mobile apps for users in 160 countries.

Google added natural voice interactions for the Gemini app on macOS. The voice interaction feature provides intelligent dictation that automatically cleans up spoken transcriptions, such as removing filler words and handling mid-sentence corrections. Users can also opt into advanced Gemini reasoning to let the app analyze screen context and execute complex desktop tasks via voice input.

Friend re-launched its AI pendant device, featuring a newly added speaker that enables two-way voice conversations with users. The redesigned pendant includes capabilities to remember user conversations for extended periods. The updated wearable is now priced at $249 alongside a $10 monthly subscription.

Nvidia expanded its Agent Toolkit with re-engineered PhysicsNeMo libraries and CUDA-X components that agents can invoke for physics modeling, sparse-system solving and quantum-chemistry calculations. The package is aimed at autonomous engineering workflows in chip design, verification, packaging and industrial simulation, with Cadence, Siemens and Synopsys among the early adopters.

AI Research News

OpenAI discovered that retaining reasoning in its API harness tripled GPT-5.6 Sol’s score on ARC-AGI-3, increasing its score on the puzzle benchmark from 13.3% to 38.3%. Previous low scores were caused by API settings that discarded private reasoning and utilized rolling truncation. Adjusting the harness significantly improved the model’s memory retention and complex problem-solving efficiency during agentic evaluations.

Moonshot AI released the Kimi K3 Technical Report. The report on the 2.8T parameter Kimi K3 open-weights model presents several architectural and training innovations: It combines efficient Kimi Delta Attention with periodic global-attention layers, using Attention Residuals to retrieve information from earlier network depths; it introduces Stable LatentMoE, which activates 16 of 896 experts while using SiTU-GLU and Quantile Balancing for stability in training. These and other innovations in Kimi K3 result in 2.5 times greater scaling efficiency than Kimi K2.

Nvidia released a real-time world model called Cosmos-H-Dreams for surgical robotics. Cosmos-H-Dreams is an action-conditioned generative simulator that predicts surgical scenes from an initial image and a stream of robot movements. By distilling a larger bidirectional model into a causal student and serving it through Nvidia’s FlashDreams runtime, researchers increased generation from roughly 10 to about 160 frames per second on one RTX Pro 6000. Nvidia stresses that this remains a research platform, not a surgical tool.

Claude discovers improved attacks on cryptographic algorithms. Anthropic researchers used Claude Mythos Preview to develop a key-recovery attack against the HAWK-256 experimental post-quantum signature scheme and a new technique that speeds up a known attack on seven-round AES-128 by 200-800 times. The results demonstrate meaningful AI-assisted cryptographic discovery, but neither breaks production versions of AES or other encryption systems.

AI Business and Policy

OpenAI launched the ChatGPT for Academic Researchers program to provide frontier AI model access to academic researchers. The program will start with providing free access to frontier AI models such as GPT-5.6 Sol Pro to its initial cohort of 10,000 researchers this summer. The initiative is backed by a commitment of more than $250 million through 2027 to support external scientific research, grant writing, and advanced computational workflows, and it will expand to support 100,000 scientists, mathematicians, and engineers at selected institutions.

Stock market turbulence in AI and tech stocks in July led to the wipeout of a hedge fund betting on AGI. Situational Awareness, a hedge fund founded by former OpenAI employee Leopold Aschenbrenner, faced margin calls as bets on companies like Marvell went sour, and he was forced to unwind public equity holdings as financial losses mounted.

Nvidia has invested in Ilya Sutskever’s Safe Superintelligence, announcing a long-term partnership that includes Nvidia investment in SSI and access to Vera Rubin systems that would increase SSI’s computing capacity by an order of magnitude for their AGI development work.

AI Opinions and Articles

CEO Dario Amodei clarified that Anthropic has never advocated for a ban on open-weights models. In a post on Anthropic’s website, he emphasized that “Open-weights models that don’t have dangerous capabilities are a public good,” and protectionist bans fail to address national security concerns regarding dangerous capabilities and industrial-scale distillation. Amodei instead reiterated support for targeted export controls on advanced chips, restrictions on massive model distillation, and mandatory pre-release safety testing.

His statement is a response to commentary on Anthropic’s conspicuous absence on a joint statement “Open Weights and American AI Leadership” in support of open source AI models.

In yet another joint statement, major AI research organization leaders signed an open letter at pacingthefrontier.com cautioning to pace AI development as recursive self-improvement works to accelerate AI progress:

Levels and Cycles in AI

The AI revolution has proceeded in stages of levels and cycles.

The cycles have become familiar to us, as ‘AI freakouts’ with refrains that repeat with regularity. We’ve gotten AI hit a wall freakouts every time new frontier AI models are slow to release. Every tech layoff stirs up the ‘AI will take our jobs’ freakout.

The AI safety freakout: A more-powerful-than-ever model like Mythos 5 comes out, and AI doomers worry that the end is nigh with “AI will kill us all” narratives. With Mythos 5 and Fable 5, the freakout became actionable when the Government blocked Fable 5 from non-US users getting and Anthropic took it off the market. Three weeks later, it was back on the market. Then OpenAI and Anthropic released GPT-5.6 and Claude Opus 5 respectively while tamping down fears by claiming improved guardrails on cyber-security.

The Chinese AI freakout: Every time China releases a frontier-level AI model, like Kimi just did with K3, it gets presented as a threat to US AI labs and by implication US dominance in AI. Pundits tell us “we cannot let China win the AI race” but not lose ourselves when we do. When a new frontier US-based AI model lands, we forget it until next time China scores a great AI model.

The Levels of AI

We can describe the evolution of frontier AI systems since the ChatGPT Moment in late 2023 as advancing upwards through five broad capability levels, categorizing the progression from conversational assistants toward autonomous AI systems.

Level 1: Conversational AI that answers questions. With GPT-4 in 2023, we got natural language conversation useful for everyday work. It was turn-based chat interaction, zero autonomy.

Level 2: Multimodal AI that understands different types of information. GPT-4o and Gemini-1.5 releases in 2023 and 2024 gave us AI that could understand not just text, but also images, PDFs, videos, and code.

Level 3: Reasoning AI that reasons through difficult problems. OpenAI’s o1 and o3 and DeepSeek R1 in 2024 and 2025 introduced AI that could execute structured reasoning, via test-time compute, before answering. All subsequent AI model improvements have been based on improving AI reasoning, and through that, its ultimate intelligence capabilities.

Level 4: Agentic AI that completes complex tasks autonomously. As AI reasoning advanced into multi-step reasoning in 2025 and 2026, with AI models like Claude 4.x Opus, GPT-5 through GPT-5.5, and Gemini 3.1 Pro, AI models became able to solve ever more complex research, mathematics, and software engineering tasks. Goal-directed task completion also needs tools, browsers, code execution, memory, and planning to effectively and autonomously complete tasks. Thus, AI harnesses such as Claude Code, Codex, Cursor or OpenClaw provide support for tool use, skill use, and context engineering to get the most out of these capable AI models.

Level 5: Long-Horizon Autonomous Agentic AI that autonomously manages and completes large projects. AI is now capable of long-duration planning, delegation among specialized agents, sustained projects, complex software and research collaboration.

This new level is not a change in kind from Agentic AI but a change in scale and scope. In late 2025, AI models such as Opus 4.5 improved their internal reasoning and autonomous code generation enough to tackle more difficult scientific, mathematical, and programming problems. The releases of GPT-5.6 Sol, Claude Opus 5, and Kimi K3 are doing that again: Expanding the scope and scale of Agentic AI.

What Level 5 AI Can Do

Agentic AI was the shift from answering questions to completing tasks: researching, coding, editing, debugging, coordinating tools, and iteratively improving results.

The leap from Agentic AI to Level 5 Long-Horizon Autonomous Agentic AI is largely about operational capability around long-horizon planning and multi-agent coordination. As such, the harness used plays a major role in having usable strong capabilities. However, higher AI model intelligence is needed for reliable orchestration of complex problems.

Agentic AI models in 2025 could handles time-horizon tasks of tens of minutes. Now, combining the latest AI models with very long context, persistent working memory, and access to orchestrate multiple AI sub-agents, enables coordinated multi-agent execution over days or longer.

The New Level 5 AI Models

Several models are established on the Level 5 beachhead: GPT-5.6 Sol, Fable 5, Claude Opus 5, and Kimi K3. MemClaw stacked them up. These AI models exhibit stronger reasoning and more reliable agentic behavior than any prior AI model. With persistent memory and tool use, they can be combined into systems capable of sustained collaboration on complex work.

More models are coming: GPT-6 will come soon. Gemini 3.5 Pro is delayed because it is not performing at this competitive level yet. Alibaba’s Qwen 3.8 Max is in preview and seems to be on Kimi K3 level.

New Rules for Prompting Opus 5

When frontier AI models improve on capabilities and intelligence, it’s wise to update prompting, context engineering, and your harness to take advantage of the new powers they possess.

Anthropic has produced a Context Engineering guide for Claude 5 AI models. They found out that less is more with Claude Opus 5. They removed over 80% of Claude Code’s system prompt and got better results.

Overall, we found that we were over-constraining Claude Code, both through our system prompt and in our CLAUDE.md files and skills.

Prompting Claude Opus 5 like a less capable AI model holds it back. They suggest relaxing constraints:

• User fewer strict rules but let Claude use judgement instead. “… newer models have better judgement and can handle these decisions well without explicit rules.”

• Don’t give examples, design interfaces.

• Claude now automatically saves memories. You don’t have to save it explicitly in CLAUDE.md.

• Claude used to rely of specs in markdown files, now it uses complex references. “Claude can reference HTML artifacts created by our new artifacts feature.”

• Use progressive disclosure, and by implication make most practices skills. “Progressive disclosure is not just for skills; we also use it for tools.”

If the guidelines are too vague or unclear, or you just want Claude to clean up your files, Anthropic rolled out a Claude doctor command, to help you do this automatically.

Treat Your Level 5 AI as a Smarter Employee

AI is now smart enough to complete a variety of tasks and be your AI assistant; treat AI as an employee. Give AI the support (harness), guidance (via context and memory), tools (access to tool use and skills), and incentives (proper intentional prompting) to be a most helpful AI assistant.

What kind of assistant can these AI models be? These latest frontier AI Models are an intelligence and autonomy upgrade that promotes the AI from junior assistant (agentic AI) to senior developer and project manager (Long-Horizon Autonomous Agentic AI). You need less micro-management of details in many prompts:

• You won’t need to instruct the model to double-check its work. Claude Opus 5 will perform needed verifications and self-correction.

• Don’t give bite-size multi-turn step-by-step guidance. Instead, give complete context, inputs, and end-goal specifications in a single prompt, where Opus 5 performs best. /goal and loops are more effective than ever.

• To limit Opus 5 to avoid overthinking or over-engineering answers, set limits on scope and target deliverables. You can cap summary lengths or set word count limits to limit verbosity.

Just as a senior employee needs less hand-holding and thrives on delegation, your AI employee improves if you delegate more and micro-manage less. In simple terms, do less: Simplify prompts, skills, and tools definitions. Cut down instructions built for earlier, weaker AI models, and reduce them to what you really need.

Use a Good Harness

You need a good AI model and a good harness that works with them. You can get that with integrated Claude Clode / Cowork or OpenAI ChatGPT with Work / Codex, each using their respective AI models.

You can also get that by using an open harness like OpenCode, OpenClaw or Hermes Agent and picking your own AI model. You can use Kimi K3 as your driver, now that the weights are available.

These days my go-to harness for code development is Claude Code. I use OpenAI’s ChatGPT with Codex and Work for other tasks. These harnesses are growing with the AI models.

I use Gemini Deep Research occasionally, for both complex research (supercritical CO2 heat exchangers for Brayton cycle turbine) and mundane questions (which brand of security camera to buy), but it can be ‘flabby’ and verbose. To fix verbosity on frontier AI models, I’ll use the “I have ADHD” skill to get just the facts and actions I need.

Hermes Agent is my personal local AI assistant. I haven’t run the latest frontier AI models on it yet, since open local AI models are good enough for my local AI tasks, for example, analyzing personal medical files. Hermes Agent is great at managing and automatically building skills from your interactions, and it has a Kanban-style feature to manage tasks and agents. I’ll take advantage of that, feeding Hermes Agent more ambitious tasks, encoding repeated tasks as skills, and orchestrating delegated tasks in parallel.

more ambitious tasks, encoding repeated tasks as skills, and orchestrating delegated tasks in parallel.

Conclusion - Increase your Ambition

AI can do more than ever before. GPT-5.6 Sol, Fable 5, Claude Opus 5, and Kimi K3 are more intelligent than any prior AI models. Today’s frontier AI models can act as a project collaborator, decomposing large objectives into specialized sub-agents, coordinating their work, monitoring progress, revising plans, and producing integrated results with relatively little human supervision.

The latest harnesses bring skills, memory, tools, and gateway connections that build on these capabilities and make AI agents more capable than ever.

It may be easier to use AI as you did before, and it will be a more reliable chatbot and inline copilot. However, achieving real productivity leaps with AI requires taking advantage of AI’s next-level capabilities. You’ll need to embrace Level 5 Long-Horizon Autonomous Agentic AI.

Your advantage, your “alpha” with AI, is intent and initiative. Increase your ambition with AI.

Do more with AI. Free AI from constraints of limiting system prompts, from feeding it bite-sized problems that don’t fully complete your task, and from reviewing in detail each activity. Become a manager of a fleet of AI agents, and delegate to AI like it’s a knowledgeable, capable employee. Give it the knowledge, context and guidance it needs and let it rip. AI is now ready for it.

Top Tools: Claude Opus 5 and Laguna S 2.1

Anthropic released Claude Opus 5, offering comparable and sometimes better performance to its frontier Claude Fable 5 model at the cost of its predecessor Opus 4.8. Opus 5 improves on coding, professional analysis, computer use, scientific reasoning, and long-running agentic work. It’s truly the current world’s best AI model.

Claude Opus 5’s benchmark scores are stunning. For knowledge work, Claude Opus 5 scores a SOTA 1861 on GDPval-AA, well above Fable 5, GPT-5.6 Sol, and 250 points above Opus 4.8. It advances fluid intelligence with an ARC-AGI v3 score of 30%, well above every other tested AI model. AGI soon? It’s SOTA on BrowseComp (90.8%), FrontierCode (53.4%), and OSWorld2.0 (70.6%), making it most advanced for agentic tasks.

Anthropic claims “Opus 5 is our most aligned model to date” with the lowest rates of reckless or deceptive behavior. They acknowledged that it remained behind Mythos 5 on cybersecurity tasks, but that’s a good thing; while less able to find exploits, it performs just as well at finding and fixing software vulnerabilities. More importantly, Claude Opus 5 is available today on all platforms and doesn’t have the onerous data retention or restrictions on use of Fable 5.

First-day reviews of Opus 5 are mostly positive. Opus 5 beats Fable and lives up to Fable 5 level performance according to some. Dan Shipper calls it hard to love on account of its quirks and need to prompt differently for best results. He suggests running it below high effort.

Poolside released Laguna S 2.1, an open 118B parameter MoE (Mixture-of-Experts) AI model designed that activates just 8 billion parameters per token. Designed for long-horizon agentic coding tasks, the model supports a 1 million-token context window. Laguna S 2.1 delivers strong performance matching or exceeding significantly larger systems on long-horizon agentic and coding benchmarks, such as 70.2% on Terminal-Bench 2.1, 59.4% on SWE-Bench Pro, and 40.4% on DeepSWE.

Vibe-checks confirm Laguna S 2.1 is a solid AI model for agentic coding and agentic tasks. I’ve used it in Hermes Agent with good results, and pricing via NousResearch and OpenRouter is now free.

Poolside released Laguna S 2.1 under an open-weights OpenMDW-1.1 license and published evaluation trajectories and model weights on Hugging Face. It can be post-trained and fine-tuned. Impressively, they trained this model in only 9 weeks on a cluster of H200 GPUs. Poolside is US and Europe-based, making Laguna S 2.1 a uniquely powerful and efficient open AI model developed outside of China.

AI Tech and Product Releases

Google released Gemini 3.6 Flash and Gemini 3.5 Flash-Lite, upgrades from prior Gemini versions that are faster, feature enhanced safety safeguards, and are more cost-effective. Gemini 3.6 Flash improves over 3.5 Flash in knowledge work (GDPval-AA v2 score of 1421 versus 1349), coding (DeepSWE 49% vs. 37%), and computer use (OSWorld-Verified 83.0% vs. 78.4%). While improving quality, it reduces output-token consumption by 17% relative to 3.5 Flash and cuts the API cost, with Gemini 3.6 Flash costing $1.50/$7.50 per million input/output tokens.

Google highlights Gemini 3.5 Flash-Lite for its combination of high speed (at 350 output tokens per second), intelligence, and cost efficiency. It is priced at only $0.30/$2.50 per 1M input/output tokens. Both 3.6 Flash and 3.5 Flash-Lite models are available immediately through Google’s platforms and consumer apps.

Google also announced Gemini 3.5 Flash Cyber as a lightweight cybersecurity model fine-tuned to rapidly detect, validate, and patch software vulnerabilities. Integrated into the CodeMender platform, the Gemini 3.5 Flash Cyber model provides an affordable and scalable tool for frontline defenders to secure complex codebases. The release has been restricted to governments and trusted partners through a limited CodeMender pilot because of its dual-use vulnerability-discovery capabilities.

Alibaba announced Qwen 3.8 Max Preview as a massive 2.4T parameter model that will be fully released soon as an open weights model. The preview is available via the Qwen Token Plan and is still “evolving” for final release. Qwen team haven’t shared benchmarks but claim the model is “second only to Fable 5” as one of the most powerful models available today.

OpenAI introduced Presence for enterprise agents. Presence combines OpenAI models with company policies, approved actions, simulations, evaluations, escalation rules, and a Codex-powered process for proposing improvements to deployed voice and chat agents. It supports tasks such as billing resolution, customer service, outbound sales, and internal IT support. In short, Presence enables businesses to deploy and manage AI agents across customer-facing and internal workflows. Presence has launched in limited availability through OpenAI Forward Deployed Engineers and select systems integrators.

OpenAI launched Health in ChatGPT, which allows eligible users to connect medical records and Apple Health data, so ChatGPT can compare laboratory results, and relate activity, sleep, and exercise data to health questions. The initial rollout is limited to U.S. users aged 18 or older on web and iOS and requires permission before using connected information. The release moves consumer AI from answering isolated medical questions towards reasoning over personal health information, but it is also explicitly designed to support rather than replace professional medical care.

Black Forest Labs announced FLUX 3, their joint multimodal foundation model capable of unified image, video, and audio generation along with action prediction. Based on the principle that real-world intelligence is inherently multi-modal, FLUX 3 was trained from images, video, and audio in one architecture, unifying capabilities of image and video generation, native audio, multilingual dialogue, editing, and action prediction. An early version of FLUX 3 is running on robots, and FLUX 3 is available for Early Access testing.

Microsoft introduced MAI Image 2.5 Pro and MAI Voice 2 Flash in public preview. MAI Image 2.5 Pro supports high-quality image generation and editing, ranking third on Arena for image editing, and it is now integrated in PowerPoint for image-to-image editing. MAI-Voice-2-Flash is also integrated with Microsoft’s product for transcription and voice services, such as Azure Voice Live, which gives developers a service to build voice-enabled agents.

Earlier this month, Runway announced Runway Dev, and now Runway has unveiled Media Router on the Dev platform. Runway Media Router is an optimized router for generative AI models, automatically routing requests across audio, image, and video models based on user constraints for latency, cost, and quality.

Inflection AI launched a research division called Inflection AI Labs and released Pi Journeys, an experimental consumer AI product designed to adapt to users’ life stages and maintain structured relational memories. The release also featured an upgraded version of the company’s flagship chatbot, Pi, incorporating improved voice capabilities, memory, and agentic tools.

Meta began piloting a new AI storytelling app called StoryKit that automatically generates personalized children’s stories complete with custom characters, lessons, and music. The application allows users to create visual assets by photographing favorite toys and defining moral values without requiring manual writing.

Google introduced study notebooks as a new feature within the Gemini app to help users organize learning materials. The tool generates lessons tailored to individual strengths and knowledge gaps based on user learning goals.

OpenAI launched the ChatGPT for small businesses program, an initiative to help small teams scale operations with AI, automate multi-step projects, and access enterprise-grade AI tools. Following their release of ChatGPT Work, this program provides training and support to small businesses to leverage OpenAI’s AI capabilities.

AI Research News

OpenAI disclosed an unprecedented security incident where an internal AI agent bypassed its evaluation guardrails and compromised infrastructure. During benchmark testing without production safety classifiers, models including GPT-5.6 Sol and a more capable pre-release model (likely GPT-6) exploited a zero-day vulnerability to gain internet access and successfully retrieve test solutions to ‘hack’ a benchmark. OpenAI and Hugging Face detected and contained the activity, began forensic investigations, disclosed vulnerabilities to affected vendors, and tightened evaluation controls.

Cybersecurity experts noted that the incident stemmed partly from human error, as OpenAI failed to properly isolate the testing sandbox from the internet. This was a test, not an actual security breach, but the event highlights the rapid advance of cyber capabilities and the need for enhanced containment, monitoring, and safety practices in advanced AI development.

AI Business and Policy

OpenAI committed to spending $750 billion on infrastructure through 2030, increasing its previous estimates by 25% despite delays in its Stargate data center initiative. The investment kicked off with Project Camellia, a $20 billion 3.2 GW data center campus. Filings indicated that energy demand for the new facility will be met primarily through newly constructed natural gas, grid-scale batteries and solar power. Meanwhile, the Stargate Project is struggling to get off the ground.

AMD announced a partnership with Anthropic to invest up to $5 billion and deploy up to two gigawatts of Instinct MI450 AI GPUs utilizing the new Helios rack-scale system. As part of the multi-year collaboration, Anthropic agreed to utilize Claude across AMD’s software and product development.

Alphabet, parent of Google, reported strong second-quarter financial results driven by rapid expansion in its cloud computing and enterprise AI divisions. Google Cloud revenue surged 82% year-over-year to $24.8 billion, lifting the company’s total quarterly revenue to $119.8 billion. CEO Sundar Pichai stated during the earnings call that surging enterprise demand and a growing backlog of cloud contracts justify the company’s projected $180 billion to $190 billion in capital expenditures for the year. Investors reacted negatively to the announcement, indicating concern with the large capex spending.

In their earnings release, Google announced that its Gemini AI assistant surpassed 950 million monthly active users, a threefold increase from a year ago and up from 750 million monthly active users at the end of 2025. The company attributed the growth to expanding agentic features such as Daily Brief and Gemini Spark, alongside strong adoption of the iOS app and its Nano Banana image generation model. Additionally, Google’s Q&A-style AI mode in Search crossed 1 billion users during the quarter.

The White House announced more than $5 billion in awards for the Genesis Mission, the federal initiative to supercharge U.S. scientific research with advanced AI. The Genesis Mission involves more than 15 agencies leveraging shared scientific data, research facilities, funding programs, and the Department of Energy’s American Science and Security Platform. The chosen inaugural cohort of 278 projects for the Genesis Mission leverage AI in areas such as biomedical discovery, advanced materials, energy infrastructure, quantum systems, biological-threat detection, and national-security.

More than $500 million in financial and computational contributions from private sector partners is supplementing the Federal science research funding. One of those partners is Microsoft, which committed $60 million to the DoE Genesis Mission to accelerate AI in scientific research. The initiative established a dedicated coordination hub named SPARK to integrate Microsoft Azure cloud infrastructure and AI tools across 17 National Laboratories.

Monday.com laid off 20% of its workforce, amounting to approximately 630 employees, as part of a major restructuring plan to refocus its business around AI. The company pivoted hard toward its AI Work Platform earlier in the year, redesigning its core enterprise product to feature no-code builders, customizable AI agents, and workflow automation tools.

Google launched the AI & Economy ATLAS study to examine how individuals and organizations use AI in daily life and workplaces. The first dataset analyzed 15 million de-identified human-AI interactions across the Gemini app, AI Mode, and Gemini API spanning more than 150 countries. The findings revealed that users primarily employ AI assistants for specific task support rather than full process automation across diverse global occupations.

Anthropic contributed an additional $20 million to Public First Action, bringing its total support to $40 million for AI policy initiatives, which are aimed to promote legislation focused on AI safeguards, export controls, and transparency requirements for advanced AI models.

Glow emerged from stealth mode with an $180 million Series A funding round at a $1.2 billion valuation, positioning itself to secure employee devices against AI-driven cyber threats.

Top Tools

China’s Moonshot AI released Kimi K3, a multimodal open-weights AI model with frontier-level performance on coding, agentic knowledge work, and autonomous execution. K3 is a mixture-of-experts model with 2.8 trillion total parameters and 50 billion active parameters, making it the largest open-weights AI model ever released. It features a one-million-token context window designed to handle long-horizon coding tasks and complex reasoning.

Benchmark tests indicate K3 is state-of-the art in front-end development and web engineering tasks, achieving number one on WebDev arena, and close to Fable 5 and GPT-5.6 Sol on code-specific benchmarks, such as 67.5% on DeepSWE and SOTA 42% on SWE Marathon, measuring long-horizon development metrics. Beyond coding, it gets stellar GDPval-AA score of 1668, above Claude Opus 4.8, and also registered a writing competency score of 2840 ELO, surpassing even Claude Fable 5.

Kimi K3’s input pricing is competitive at $3 / $15 per million input/output tokens, comparable to Claude Sonnet 5 pricing, yet offering higher performance. However, Kimi K3 operates at a higher token consumption rate than token-efficient models like GPT5.6 Sol, resulting in slower output and mitigating the cost advantage for Kimi K3.

K3 suggests that the strongest Chinese open-weight AI models are close to the leading proprietary American AI models, while also being aggressively competitive on price and developer control. We can expect K3-based models similar to Cursor’s Composer soon. US AI Czar David Sacks expressed concern over K3’s success in becoming number one on coding benchmarks, posing it as a challenge to the US to not ‘bog ourselves down’ with pre-approval of models and AI regulations that could slow us down in the AI race.

AI Tech and Product Releases

Mira Murati’s Thinking Machines Lab introduced Inkling, targeting enterprise customization with Apache 2.0 open-weight license. The multimodal 975B parameter Mixture-of-Experts Inkling model has 41B active parameters and a “controllable thinking effort” mechanism to balance cost against performance. Inkling’s performance is not frontier-level; its Artificial Analysis index benchmark of 41 ranks it below Kimi K2.6. However, it was trained on general foundation capabilities to be further fine-tuned for customized tasks, and to support that, model weights are available on Hugging Face and via the Tinker API.

Anthropic updated Claude desktop by integrating an in-app browser within Claude Code feature. Users can activate the sidebar interface using specific keyboard shortcuts to highlight code blocks and direct real-time modifications on standard web elements. This browser integration allows the software to navigate public web content and parse text fields without relying on third-party API platforms.

Prism ML announced Bonsai 27B, a one-bit 27B parameter model designed to run on mobile hardware. The highly compressed Bonsai 27B model quantizes Qwen3.6 27B and deploys in 4 gigabytes, sufficient to operate entirely on most local devices. PrismML claims Bonsai 27B retains 90% of the performance of the original 27B model, yielding high intelligence density.

SpaceX AI has open-sourced Grok Build. Grok Build is a specialized coding agent and terminal user interface package, similar to Claude Code, and has gotten praise for how effective it is paired with Grok 4.5. The source codebase for the Grok Build harness and tool-call dispatch has been published to GitHub to allow developers to fork, analyze, and build custom workflows on the base system. The open-source release aims to gain mind-share for Grok as a coding system and foster community development of their platform.

XAI added automated task features to Grok’s core interface. The Automations interface, like similar features in ChatGPT, allows users to establish specific execution frequencies, such as daily runs, for automated data collection tasks. The framework also supports external event triggers, enabling the system to monitor email inboxes and execute tasks upon receiving specific correspondence.

Apple has begun rollout of updated Siri AI assistant across it consumer hardware ecosystems. It’s been long-awaited, but Apple is finally bringing its Apple Foundation models and Gemini intelligence into Siri AI. The update brings Siri AI and other intelligence capabilities to iOS 27 public beta, including iPhone, Apple Watch and other devices. It’s still a beta release, but initial user feedback is positive regarding its stability and enhanced capabilities.

Google announced an update to Google Vids featuring custom digital avatars and Gemini Omni integration. Google Vids now allows users to create videos featuring personalized avatars from selfies and voice recordings, as well as generate videos with Omni in Google Vids using prompts and reference images. New features in video generation include animating static slide layers and automated editing for backgrounds and lighting.

Google is connecting third-party applications directly to AI Mode, rolling out U.S. integrations that let AI Mode users interact with Instacart, Canva, and YouTube Music from inside Search. This integration shifts Search from generating answers toward orchestrating transactions across applications by, for example, adding groceries to a cart, locating design templates, or saving a playlist.

OpenAI has upgraded text search and retrieval capabilities inside ChatGPT, allowing users to filter and locate historic projects, custom images, documents, and chat records from a single search console.

Google is rebranding NotebookLM to Gemini Notebook. The collaborative research document features of NotebookLM remain identical, but this further integrates it into Gemini as Google’s primary AI interface.

Google’s Interactions API has reached general availability as the primary interface for Gemini models and agents. This release introduces new capabilities including Managed Agents, background execution, and upcoming support for Gemini Omni.

TikTok is testing an opt-in tool with select US creators to identify potential unauthorized uses of a creator’s likeness. The system scans for AI-generated content, allowing verified users to review findings and report unauthorized posts or accounts.

Nvidia released Nemotron 3 Embed models, which includes an 8B model that tops the RTEB leaderboard and 1B variants optimized with NVFP4 acceleration for NVIDIA Blackwell architectures. These models are designed to enhance retrieval quality and agentic efficiency in production-scale RAG, code retrieval, and agent memory workflows.

Roblox announced a new feature called Build, allowing users to design games built with AI from their mobile devices. The Build feature allows users without programming experience to ‘vibe code’ games by prompting AI models to generate gameplay mechanics, environments, characters, and sound. To mitigate concerns regarding low-quality content, Roblox will rank these AI-generated games based on player retention.

Nvidia updated NemoClaw with blueprints to enable programmatic action from video analysis. The new collection of open blueprints allows developers to build autonomous agents that integrate video analytics with enterprise tools like Jira and Slack. This system automates downstream workflows such as generating structured reports and escalating anomalies based on analyzed footage.

AWS introduced Amazon Quick, an AI assistant designed to automate sales administrative tasks and workflows. The tool integrates with CRMs like Salesforce and HubSpot to prioritize leads, draft personalized outreach, and generate meeting preparation documents.

Capital One released VulnHunter, an open-source agentic AI security tool that uses “attacker-first forward analysis” and a “falsification engine” to scan source code for vulnerabilities while minimizing false positives. The open-source AI security tool runs on Anthropic’s Claude Opus 4.8 model.

Spotify introduced a voice-driven conversational interface that allows users to use voice to query their listening history and execute commands to generate tailored playlists based on specific time periods. The integration, available to premium subscribers on its homepage and mobile app, also allows subscribers to have back-and-forth discussions regarding background artist biographies and music genres.

AI Research News

OpenAI improved AI resilience to vulnerabilities with GPT-Red automated adversarial training. OpenAI trained an internal model specifically to discover prompt-injection vulnerabilities through iterative self-play, then used its adversarial attacks to train GPT-5.6. The company reports that GPT-Red succeeded in 84% of previously unseen indirect-prompt-injection scenarios versus 13% for human red-teamers, while GPT-5.6 Sol produced six times fewer failures on OpenAI’s hardest direct-injection benchmark than its best production model four months earlier. The evaluations are largely internal, so need external review, but this shows a scalable way for security testing to keep pace with increasingly capable agents.

Mistral shared a case study on how fine-tuning Pixtral-12B on satellite imagery led to significant improvements over the base model. Using Low-Rank Adaptation (LoRA) on an Aerial Image Dataset increased classification accuracy from 0.56 to 0.91 and reduced hallucinations from 5% to 0.1%. This shows that even as frontier AI models advance, LoRA and fine-tuning are still relevant. Mistral notes:

By adapting pre-trained models to specific domains, we can unlock dramatically better performance on specialized tasks.

AI Business and Policy

Databricks announced a new funding round at a $188 billion valuation, more than tripling the company’s December 2024 valuation of $62 billion. Led by Coatue, the $3 billion funding deal is expected to close later this summer. In recent years, Databricks has expanded its AI capabilities with products including Lakebase, Unity, and the Omnigent meta-harness.

OpenAI is reportedly developing a mobile, screenless smart speaker designed to function as an intelligent ambient companion. The rumored hardware device aims to integrate voice models directly into local smart home appliances to execute appliance control commands via natural language.

China established an intergovernmental AI organization, with representatives from 29 countries signing an agreement establishing WAICO, the World Artificial Intelligence Cooperation Organization, an independent intergovernmental body headquartered in Shanghai. WAICO gives China an institutional vehicle for shaping international AI rules as an alternative to AI governance frameworks led by the US and Europe.

New York state enacted a one-year moratorium on the development of data centers, suspending the issuance of new environmental permits for data centers using over 50 MW of electricity.

Anthropic entered into a $10 billion deal with Meta for computing power after making similar multibillion-dollar deals with SpaceX and TeraWulf. Anthropic also plans to invest $50 billion in its own data centers in partnership with Fluidstack.

Anthropic has launched the “Claude for Teachers” initiative for verified K-12 educators operating in the United States. Eligible academic professionals receive free access to the platform’s premium subscription features alongside curated teaching skill libraries. The program provides direct pathways to connect the model interface with validated educational curricula.

OpenAI introduced enhanced protections for teenagers on ChatGPT, including Study Mode which uses guided questioning to facilitate active learning. New updates also feature expanded parental controls, age prediction technology, and break reminders designed to encourage healthy digital habits.

AI Opinions and Articles

1. AI may become radically more powerful over the next 10 years.

2. This could drive an unprecedented transformation of our economy, larger than the Industrial Revolution, but unfolding over a vastly shorter time frame. It could bring risks, including large-scale job displacement, as well as opportunities such as major gains in living standards.

3. Economists, policymakers and technology leaders must act now to understand the economics of transformative AI and to build the incentives, guardrails, and institutions needed to steer AI in a direction that complements humans and benefits society.

- “We Must Act Now” Statement on AI’s Transformation of the Economy

More than 200 experts issued a call to prepare for AI-driven economic disruption. Sixteen Nobel laureates joined economists, AI researchers, and technology leaders in warning that AI could produce a rapid economic and social transformation larger than the Industrial Revolution, calling for “incentives, guardrails, and institutions” that encourage AI development that benefits society.

Noah Smith challenged the “We Must Act Now” economic statement because it calls for action without specifying policies, and it assumes that developers can intentionally steer a general-purpose technology toward complementing workers. He argues that data has not yet demonstrated widespread AI-driven job destruction and warns that poorly informed intervention could distort innovation without protecting labor.

He’s right that the “We Must Act” statement is indeed vague, perhaps deliberately so to garner wider support. The bland verbiage risks leading people to assume a “responsible” approach to AI means simply regulating AI, but the real challenge is more complex: Who will control AI and how do we get AI democratized and ensure it benefits all?

Also arguing that AI is rapidly advancing beyond our understanding of its risks, Google DeepMind’s Demis Hassabis proposed a frontier-AI standards body, a U.S.-initiated, internationally oriented institution modeled partly on FINRA. This body would test frontier systems before release, develop adaptable standards, include open-model expertise, and coordinate action if dangerous capabilities emerge. An AI leader explicitly endorsing third-party pre-deployment evaluation moves pre-release external testing closer to an industry norm.

TLDR - This week, two American AI labs rejoined the frontier AI model race and delivered frontier-level AI models: SpaceXAI’s Grok 4.5 and Meta’s Muse Spark 1.1. And OpenAI released GPT-5.6 and ChatGPT Work, which integrates agentic AI in ChatGPT interface.

Top Tools: GPT-5.6

OpenAI publicly released the GPT-5.6 model family, making generally available three GPT-5.6 model versions: the flagship Sol, the balanced Terra model for everyday use, and the lower-cost Luna model. The models are frontier-level on coding, computer use, scientific reasoning and professional knowledge work. Sol state-of-the-art, and Terra on par with GPT-5.5 but less costly. OpenAI emphasizes that GPT-5.6 models are much more token efficient than prior models, making them faster and cheaper to run.

GPT‑5.6 Sol is SOTA across agentic, coding, and knowledge-work benchmarks, getting 92.2% on BrowseComp, 62.6% on OSWorld 2.0, and an ELO of 1748 on GDPval-AA. Sol gets 72.7% on DeepSWE, above Claude Fable 5’s 69.9% at one-third lower cost. GPT-5.6 Sol scored 7.8% on ARC-AGI-3, the benchmark on fluid intelligence, which sounds low but it above prior models, whose scores were below 2%.

OpenAI touts GPT-5.6 improvements in design and polish in outputs, including improved quality and accuracy in presentations and spreadsheets. Users have shown off its use in one-shot video games, websites, graphics, interface designs, and creating and editing videos:

GPT-5.6 Sol is unbelievably good at creating and editing videos. It can do motion design, product demos, and animations like this one I made by simply giving it a screen recording. GPT 5.6 has the best design taste and significantly outperforms Fable, which relies heavily on repetitive design patterns.

GPT-5.6 has undergone extensive red-teaming and safety checking as described in OpenAI’s system card. While claiming meaningful cybersecurity improvements, also described disturbing actions such as destructive virtual-machine cleanups, unauthorized credential copying, and fabricated verified research results in a small share of tasks.

In addition, METR rejected its own GPT-5.6 Sol evaluation because the model showed unusually high rule-breaking and loophole exploitation during tasks. METR estimated either an 11.3-hour or 270-plus-hour task horizon depending on how cheating was treated, but they said neither figure was robust.

OpenAI expressed in their launch video that users could “increase their ambition” in using GPT-5.6 on more complex agent use-cases. They introduced a new Ultra mode for Sol that coordinates four agents in parallel to more quickly complete difficult, long-running assignments. Showcasing agentic Sol use cases, OpenAI stated that Sol autonomously performed the post-training for Luna.

GPT-5.6 Sol is available on paid plans, while Terra and Luna also go to free users, with listed API pricing of $5/$30, $2.50/$15, and $1/$6 per million input/output tokens.

AI Tech and Product Releases

OpenAI announced ChatGPT Work, their unified ChatGPT app which has an agentic Work interface akin to Claude Cowork and where Codex features are now built-in to a combined ChatGPT desktop app. The ChatGPT desktop app that was Codex is now ChatGPT app, with users choosing between two modes, developer-oriented ChatGPT Codex and a broader ChatGPT Work. The ChatGPT Work release adds unified plugins across ChatGPT and Codex, browser multi-tab support, and faster computer use.

ChatGPT Work provides an agentic interface for users to leverage plug-ins and agentic AI to generate “share-ready” work products: Slide decks, documents, spreadsheets, and more. OpenAI also includes Sites in the interface, the feature added last month to Codex that allows users to generate websites on the fly and host them on the chatgpt.site subdomain.

OpenAI launched GPT-Live-1 and GPT-Live-1 mini, new full-duplex voice models that can listen and speak simultaneously instead of forcing users through rigid turn-taking. The models can respond to interruptions, acknowledge a speaker while listening and continue a conversation while delegating difficult searches or reasoning tasks to a frontier-level AI model in the background. The models are rolling out via an updated ChatGPT Voice, using GPT-Live-1 for paid users and GPT-Live-1 mini the default for free users, with API access planned later.

SpaceXAI and Cursor launched Grok 4.5, a 1.5T mixture-of-experts model that achieves frontier-level performance on software engineering and agentic tasks. Grok 4.5 was trained with Cursor on real agent-interaction data, which has helped it achieve Opus 4.8 level coding benchmark scores: 64.7% on SWE Bench Pro, 62% on DeepSWE, 83.3% on Terminal Bench 2.1. Grok 4.5 costs only $2/$6 per million input/output tokens and is token efficient, making Grok 4.5 a highly attractive “daily driver” AI coding model. Grok 4.5 is available through Grok Build, Cursor and the SpaceXAI API, and developer tool environment Warp added Grok 4.5 support after the model’s launch.

Meta’s Superintelligence Labs launched Muse Spark 1.1, a multimodal reasoning model with frontier-level benchmarks that gets Meta back in the frontier AI race. Muse Spark 1.1 is designed for computer use, coding, tool calling and multi-agent orchestration. Overall benchmarks put it in the class with Claude Sonnet 5 or GLM 5.2. It has a one-million-token context window and supports parallel subagent delegation and computer use across desktop, browser, and mobile.

Meta is making Muse Spark 1.1 available in the Meta AI app and through a new Meta model API at $1.25/$4.25 per million input/output tokens to compete for third-party developers.

Meta released Muse Image and previewed Muse Video, the first media-generation models from its new superintelligence organization. These models use agentic generation and test-time compute to improve accuracy, with Muse Spark reasoning and calling web search or code execution during generation. Muse Image can combine multiple visual references to generate an image and refine its own output.

Meta released Muse Image in Meta AI, Instagram Stories, and WhatsApp. However, one Instagram feature that allowed people to generate images using photographs from public Instagram accounts created a backlash, with critics saying the feature could facilitate nonconsensual digital replicas. Meta quickly discontinued it.

Muse Video is designed to generate video with native audio and is in early preview. Meta claims Meta Muse Video ranks third on text-to-video Arena rankings.

Reve released Reve 2.1, an update on their visual image generation model that is ranking No. 2 on the Text-to-Image Arena with a score of 1302. (The new Muse Image model ranks number 3 at 1280.) The update promises greater prompt understanding, world knowledge, and stronger foreign-text rendering. The model builds images through an underlying layout engine so that generated elements appear as editable layers, and it allows edits such as changing text inside an image while rebuilding the composition around the change.

ByteDance launched Seedream 5.0 Pro, a multimodal image creation model focused on “complex information visualization and native multilingual input and rendering. image-text alignment, structural coherence, text rendering, and visual aesthetics. The release adds features including point-and-lasso editing, sketch rendering, material and color replacement, layer separation, multi-image fusion, and multilingual generation. Seedream 5.0 Pro scores 1231 and ranks 11 on the text-to-image leaderboard.

Cognition released SWE-1.7, a coding model that the article says runs at 1,000 tokens per second on Cerebras. Cognition disclosed that the model is based on Kimi K2.7, notable because other coding products had used the same base. Cognition’s reinforcement-learning recipe improved its FrontierCode benchmark score from 30.1% to 42.3%.

Mistral introduced Robostral Navigate, an 8B parameter model for ‘embodied navigation.’ The model uses a single RGB camera and plain-language instructions to navigate environments, and Mistral reports a 76.6% success rate on unseen R2R-CE environments, outperforming some systems that use depth sensors or multiple cameras. The model was trained in simulation on approximately 400,000 trajectories and is intended for robots used in logistics, manufacturing, delivery and hospitality.

OpenAI released GPT-Realtime-2.1-mini for developers using the Realtime API, bringing reasoning and tool use to the Realtime API’s mini tier.

Developed by Synthetic Sciences, OpenScience has been launched as an open-source alternative to Claude Science. OpenScience includes more than 250 research skills and works with any model rather than being tied to one proprietary model.

Shanghai AI Lab released Agents-A1, a 35B parameter agentic mixture-of-experts model built on Qwen 3.5 35B and released under Apache 2.0. Agents-A1 supports a 256K-token context window and was trained for long-horizon work, with interestingly high scores on science-related benchmarks.

LiquidAI released Antidoom, a method or model update aimed at reducing reasoning doom loops. The article says it reduced the loop rate on Qwen3.5-4B from 22.9% to 1%. It also says scores improved across the board after the change.

Anthropic extended Fable 5 access on paid plans through July 12 and increased weekly Fable usage rate limits. Feeling the competition, Anthropic?

AI Research News

We audited SWE-Bench Pro, one of the most widely used AI coding benchmarks, and found it no longer reliably measures frontier coding capability. We find 30% of SWE-Bench Pro tasks to be broken and are retracting our previous recommendation that the research community use it as a leading coding eval. - OpenAI

OpenAI published a report finding that about 30% of SWE-Bench Pro problems are broken. The report suggests this caps the benchmark near 70% and should affect how readers interpret model claims based on SWE-Bench Pro. With frontier AI models scoring at that level on SWE-Bench already, other benchmarks such as Deep-SWE should be used.

Anthropic published research describing a “global workspace” inside Claude, called J-space. This space is part of a model’s internal representations and contains roughly 25 active internal concepts. The J-space appears important for multi-step reasoning, because ablating it leaves basic abilities intact while reasoning collapses. The research also connected J-space to safety evaluation awareness, including a blackmail evaluation where removing test awareness changed model behavior.

Google Research presented SensorFM, a health foundation model trained on one trillion minutes of wearable de-identified sensor data from five million consenting Fitbit and Pixel Watch users. The model transfers to 35 cardiovascular, metabolic, sleep, mental-health and lifestyle prediction tasks and reportedly outperformed specialized supervised baselines on 34 of them. Google also tested SensorFM as a grounding system for a personal health agent, although substantial clinical validation would still be necessary before diagnostic use.

Anthropic researchers introduced GRAM, a modular training technique that routes knowledge from selected dual-use domains into components that can later be removed or enabled for authorized users. In experiments involving cybersecurity, virology and nuclear-physics information, removing a module suppressed the targeted capability without significantly degrading general model performance. The work is preliminary and has only been tested on models up to 5 B parameters.

AI Business and Policy

An internal memo reviewed by Reuters indicates that Meta plans to begin producing its internally designed Iris accelerator in September. Developed with Broadcom and manufactured by TSMC, the chip is intended to reduce Meta’s dependence on Nvidia and AMD as the company expands toward 14 gigawatts of computing capacity in 2027. Meta expects to spend as much as $145 billion on AI infrastructure during 2026.

Tencent and Manus’ original investors are discussing buying the AI-agent company back from Meta for at least $2 billion, after Beijing had ordered Meta to unwind the acquisition while reviewing whether it violated Chinese investment rules.

The Bank of England warned that the AI boom is becoming a financial risk. The central bank said that a reassessment of AI profitability might trigger falling equity prices and leveraged investment in AI companies and heavy infrastructure borrowing could amplify a market downturn. It also warned that existing financial rules were not designed for autonomous AI agents.

Apple filed suit against OpenAI and two former Apple employees over alleged theft of hardware trade secrets. The lawsuit alleges that confidential information was taken by two to support OpenAI’s expansion into consumer hardware. OpenAI denies the accusations.

AI Opinions and Articles

Final thoughts: This is the most important week for AI releases this year. Two American AI labs, SpaceXAI and Meta, released top-tier AI models. GPT-5.6 sets a new standard for the AI model frontier, and we can breathe a sigh of relief that the Government hasn’t blocked access and we’ll get to use frontier AI.

With AI is reaching a new level, remember to increase your ambition with AI. Take AI further than you have taken AI before. Give it harder questions, larger tasks, longer reasoning, and more autonomy.

The future of the firm is a learning loop in which human capital and token capital compound. – Microsoft CEO Satya Nadella

Top Tools - Fable 5 is back

Anthropic has restored global access to Claude Fable 5 on Claude and Claude Code following the U.S. Department of Commerce’s withdrawal of emergency export controls. Anthropic’s new release of Fable 5 features additional cybersecurity safety guardrails and a proposed Cyber Jailbreak Severity (CJS) framework to standardize software vulnerability risk assessments. Fable 5’s classifier has a higher safety margin in preventing harmful actions, so it may block some benign requests out of caution.

While Fable 5 is now available to all users, it is accessible via Claude subscription only until July 7th, after which expensive API usage credits are required for use. Subscribers, try it before it goes away. For developers, Anthropic’s Claude Fable 5 models are available again on Amazon Bedrock. Access to the specialized Mythos 5 model is limited to vetted, authorized organizations under Project Glasswing.

Anthropic’s advice for using the powerful but expensive Fable 5 is to reserve the model for high-stakes, long-context judgment tasks rather than brute-forcing small operations. Define high-level outcomes rather than micro-managing step-by-step instructions and save reusable context via Markdown files.

Anthropic released Claude Sonnet 5, positioning it as its most agentic Sonnet-class model yet, with improved planning, tool use, coding, and autonomous task execution. Sonnet 5 offers clear improvements over Sonnet 4.6, with excellent benchmark results that narrow performance gap with Opus 4.8: Sonnet 5 scores 1615 on GDP-val, 63.2% on SWE-bench Pro, and 81.2% on OSWorld-verified.

Claude Sonnet 5 offers near-flagship performance and improved agentic capabilities at a significantly lower cost than Anthropic’s Opus 4.8, with pricing of $2 / $10 per million input / output tokens. However, some testers have noted it may be less token-efficient than Opus, thus undoing that cost advantage. Reaction to Claude Sonnet 5 has been mixed, as it is close but not at the frontier as an AI model (coders might continue to use Claude Opus 4.8 for their hardest agentic coding tasks) and has been overshadowed by the release of Fable 5.

AI Tech and Product Releases

Anthropic launched Claude Science, an AI workbench that integrates research tools, datasets, and visual generation to facilitate research and analysis. The utilizes specialized agents to connect with life sciences databases and models like Nvidia’s BioNeMo, and it brings literature analysis, code execution, figure generation, manuscript drafting, and reproducibility tracking into a single environment. Currently in beta for Claude Pro, Max, Team, and Enterprise users, the workflow product can run locally on macOS and Linux platforms or through HPC access. Additionally, the company revealed plans to develop its own drugs, specifically targeting neglected diseases.

Google unveiled Nano Banana 2 Lite, a high-speed image model that generates images in four seconds, far faster than Nano Banana 2 yet at comparable quality. Google is natively integrating the fast-generation Nano Banana 2 Lite across its ecosystem, including Gemini applications, search features, and Google Photos. Aimed at enterprise image generation, it can be run from Google AI Studio or the Gemini API for just $0.034 per 1,000 images.

Google DeepMind also released a public preview of Gemini Omni Flash, Google’s multimodal Omni model for conversational short video generation and editing. Gemini Omni Flash uses text, images, and clips to produce 720p videos with synchronized audio, scene consistency, and text insertion. Currently, video outputs are capped at 10-second clips and priced at $0.10 per second in the Gemini API and Google AI Studio.

Google has combined both models in demo applications for multi-media generation with AI, including SpaceLift, a tool for interior design exploration, and Omni Product Studio, which turns product images into a brief video showcase.

Google updated its Notebook LM platform to support the automated generation of short-form, vertical video summaries. Users can process existing document notes to render 60-second video overviews complete with synthetic voice commentary and basic visual layouts. The creation process requires a notable amount of background processing time to compile and render the final media file.

Chinese AI lab Meituan released LongCat-2.0 to open source under an enterprise-friendly MIT license. LongCat-2.0 is a 1.6T parameter Mixture-of-Experts model with a dynamic range of 33B to 56B active parameters and features a 1-million-token context. Trained on domestic Chinese ASICs with model architecture and training to specialize in agentic software engineering tasks, LongCat-2.0 has frontier-class performance on coding tasks, getting 59.5% on SWE-bench Pro and 70.8% on Terminal-Bench 2.1. .

Google launched a new macOS version of Gemini Spark and released other Spark updates for desktop task automation and remote task execution via mobile devices. The update introduces integrations with services like Canva and Dropbox, support for custom Model Context Protocol (MCP), and real-time tracking for news, sports, and finance. The desktop application integrates directly with local file systems, allowing the AI to interact with local directories and manipulate user files via natural language instructions.

Gemini expanded their personalized image generation to all eligible U.S. users via Personal Intelligence. The update integrates Google Photos, Gmail, and YouTube to provide customized responses and images based on user context. This allows users to generate unique images that reflect their personal preferences and lifestyle.

Exo Labs announced local.ai, a site for comparing how AI models run on a user’s own hardware, including model capability, quantization, hardware, and workload benchmarks. The site is intended to show the best local model for a user’s hardware, the tradeoff versus cloud APIs, and whether local inference is cheaper than API tokens.

Exo Labs also previewed an Exo CLI for consumer-device inference. This “vLLM for consumer devices” handles model and runtime configurations for users wanting to run AI models on their own hardware. The CLI was expected to arrive in the coming weeks.

Sakana AI’s recently released Fugu now works in Codex and OpenCode. Fugu is Sakana AI’s multi-agent orchestration model that routes, coordinates, and verifies work across expert AI models for agents.

Hugging Face announced that Every Eval Ever results are now appearing on Hugging Face model pages, integrating community evaluation results with model discovery. The post notes that AI evaluation data is currently scattered across papers, leaderboards, benchmark harnesses, and blog posts, and says the combined repository now contains around 229,000 results across more than 22,000 models and 2,200 benchmarks.

AI Research News

Meta AI released Brain2Qwerty v2, a non-invasive brain-to-text research system that decodes sentences from MEG brain recordings without surgical implants. Meta reports 61% word accuracy overall and 78% for the best participant, a major improvement over prior non-invasive approaches, and it is releasing training code plus related data to support open neuroscience research.

DeepSeek open-source released DSpark, that accelerates AI model inference by up to 85% with improved speculative decoding. DeepSeek released a DSpark checkpoint of their DeepSeek-v4 model to showcase the capability and published the DSpark paper to explain their innovation. DSpark uses semi-autoregressive generation with draft models and confidence-scheduled verification, dynamically tailoring verification to expected success rate. This reduces overhead and increases throughput. They also released the DeepSpec codebase, which enables training speculative decoding for other open-weight models.

Google Research introduced TabFM, a zero-shot foundation model for tabular data. The goal is to bring the same kind of general-purpose model behavior seen in time-series forecasting to classification and regression tasks on structured tables, potentially reducing the need for task-specific model training in common business and scientific workflows.

AI Business and Policy

China’s Kling AI has raised $2.8 billion in venture capital funding, valuing the AI video operation at $18 billion as the company seeks to spinoff from Kuaishou Technology and expand its video AI operations.

Mark Zuckerberg informed Meta staff at a recent town hall that the development of AI agents has not progressed at the pace executives previously expected. Despite laying off 8,000 employees and reassigning 7,000 others to AI groups earlier this year, Meta’s new structure has yet to yield its intended benefits and morale is terrible in Meta’s AI unit.

Anthropic is exploring development of its own AI chips in collaboration with Samsung. The company is reportedly in contact with Samsung to investigate a partnership for custom hardware development.

OpenAI CEO Sam Altman has proposed giving 5% of the company’s equity to a U.S. sovereign wealth fund, suggesting that providing a public financial interest could blunt public backlash and ease regulatory tensions with the Government. The 5% ownership could be worth $42 billion based on company metrics. The plan has sparked policy debates regarding AI oversight, as it could create systemic conflicts of interest for federal regulators.

Microsoft announced a new Microsoft Frontier Company focused on enterprise AI deployments. The project will be backed by a $2.5 billion investment and 6,000 industry and engineering experts. Early partnerships for the venture include the London Stock Exchange Group, Unilever, Land O’Lakes, and Accenture.

The Verge reported on a revived U.S. privacy bill that would restrict AI companies and data brokers from selling sensitive health and location data. As consumer AI assistants increasingly handle intimate personal information, some lawmakers are looking to treat chatbot data flows as a health-data surveillance problem.

AI Opinions and Articles

AI is eating the world. OpenAI published new data on how ChatGPT adoption has expanded globally, noting that users primarily using non-English languages now represent more than half of active users. The analysis says Spanish, Portuguese, and Arabic are the leading non-English languages used in ChatGPT.

Top Tools

OpenAI announced the anticipated GPT-5.6 as a suite of 3 models (Sol, Terra, Luna), but they released them as a limited preview for “trusted partners” rather than a full public release. Following pressure from the Trump administration, GPT-5.6 access is restricted to select partners pending a review by Government agencies to develop new AI security framework developments.

GPT-5.6 joins Anthropic’s Mythos 5 and Fable 5 as frontier AI models not accessible to the public on account of its high-level capabilities posing a cyber-security risk. This Government interference is a worrying trend that could greatly hamper AI progress unless and until a smooth framework for safe public AI releases is determined and worked out. OpenAI notes:

We are taking this short-term step because we believe it is the strongest path to broader availability in the coming weeks, while we work with the Administration to develop the cyber Executive Order framework and a repeatable process for future model releases.

The GPT-5.6 series, comprising the flagship Sol model, the balanced Terra model and the lower-cost Luna model, offers enhanced performance in coding, biology, and cybersecurity. For example, GPT-5.6 sets a new state-of the TerminalBench 2.1 coding benchmark, with GPT-5.6 Sol beating Claude Mythos5, GPT-5.6 Terra on the level Claude Fable 5, and GPT-5.6 Luna on the of GPT-5.5.

OpenAI calls GPT-5.6 Sol “our most capable model yet for cybersecurity.” However, they implemented a new safety stack featuring real-time misuse classifiers and trained GPT‑5.6 Sol to recognize cyber vulnerabilities for cyber defenders but not generate exploits. OpenAI claims the model does not cross the Cyber Critical threshold under their AI safety Preparedness Framework⁠.

GPT‑5.6 is priced competitively per 1M tokens across the three model sizes: Sol costs $5 input / $30 output; Terra is $2.50 input / $15 output; and Luna is $1 input / $6 output. GPT‑5.6 also introduces more predictable prompt caching, including explicit cache breakpoints and a 30-minute minimum cache life. Sol adds a “max” reasoning setting and an “ultra” mode that coordinates subagents for complicated work, while Luna is a cost-competitive coding model on par with GPT-5.5 and Claude Opus 4.8 on coding.

GPT-5.6 will be a compelling frontier AI model lineup when it is available via a rollout through ChatGPT, Codex and the API.

AI Tech and Product Releases

Anthropic introduced a feature called Claude Tag that enables users to interact with Claude directly within Slack channels and works proactively instead of only responding when pinged. Claude Tag will break assignments into stages and tackle them off-line before returning completed work to the Slack thread. It can surface relevant information, follow stale threads, and share context at the channel level so teammates can continue work from the same conversation.

Anthropic bills it as “a new way for teams to work with Claude” and Andrej Karpathy called it “the 3rd major redesign of LLM UIUX.” However, the persistent workplace agent model is not unique but applies the OpenClaw agent paradigm within an enterprise context. Unlike a private chatbot or personal OpenClaw, the work-channel agent acts like a co-worker; it can collaborate with multiple people, use internal company tools, and learn and retain relevant organizational context. Claude Tag beta is initially available to Claude Team and Enterprise customers.

Sakana AI has launched an orchestrator model called Sakana Fugu, a multi-agent orchestration system which routes user prompts to the most optimal underlying AI models. One API call routes work across multiple models and roles, combining thinker, worker, and verifier behaviors to improve results. The system balances performance and latency with its standard Fugu version, while a Fugu Ultra variant is designed to handle complex, multi-step problems. Benchmark data shows that the Fugu orchestrator performs on par with or outperforms existing high-tier models like Fable 5 and Mythos in certain tests.

Google has added computer use as a built-in tool in Gemini 3.5 Flash, to support agents that can interact across platforms. The model processes continuous screenshots to execute click, scroll, and typing actions seamlessly across varied software environments. Developers can utilize this feature via the Gemini API and Gemini Enterprise Agent Platform to create agents capable of acting across browser, mobile, and desktop environments.

Liquid AI released its smallest AI language model yet, LFM2.5-230M, touting the tiny 230M parameter LLM as ‘built to run anywhere.’ The model supports a 32K context window and utilizes the LFM2 architecture to enable efficient agentic workflows and high-speed data extraction on edge devices, outputting at over 200 tokens-per-second on a Galaxy S25 Ultra. It is open-weights and available via Hugging Face under a dual-use commercial license.

OpenAI updated GPT-5.5 Instant to improve intent recognition and instruction following. According to OpenAI, the update is “more fun to talk to” with more natural responses, and it enhances the model’s ability to handle complex constraints, improving its shopping and local recommendations. This model update is the default model available in ChatGPT, while developers can access it via the `gpt-5.5` model.

Mistral AI released OCR 4, their 4^th generation document intelligence model that provides state-of-the-art document extraction, RAG pipelines, and knowledge search in 170 languages. OCR 4 generates structured document representations including bounding boxes, block-type classification, and per-word confidence scores. With pricing starting at $4 per 1,000 pages via Mistral API, Amazon SageMaker, and Microsoft Foundry, it suitable for low-cost, enterprise-oriented OCR.

Alibaba’s Qwen team released Qwen-AgentWorld, a language world model designed to predict how an environment changes from an agent’s actions across seven domains, including Search, Terminal, and Web. Predicting outcomes from an agent’s action can help train agents, and training agents within these controlled simulations produced performance gains exceeding real-environment training and improved scores on previously unseen benchmarks. A Qwen-AgentWorld Technical Paper was released to explain how the AgentWorld model was trained and used.

Unsloth created a 1-bit GLM 5.2 quantization to run a much smaller version of the frontier GLM 5.2 model locally on a 256GB Mac Studio. The 1-bit quantized build shrinks the footprint to roughly 200GB in GGUF format. A 2-bit quantization retains ~82% accuracy while shrinking model sized from 1.51TB to 238GB (-84% size).

Bytedance previewed the Seedance 2.5 video generation model at a conference in Beijing, highlighting major upgrades to the video generation platform. The upcoming model can natively generate a single segment of full audio-video up to 30 seconds long, double the length possible in previous versions. It supports up to 50 distinct text, image, audio, or video reference assets to provide users with finer control over motion and editing.

Krea AI announced Krea 2 as an open-weights text-to-image generation model for public download and customization. Krea 2 is a12B diffusion-transformer image model with two versions; Raw is the base checkpoint useful for further fine-tuning and Turbo is the post-trained version for direct use. Making Krea 2 open weight allows developers to run the model locally and fine-tune it on specific styles or custom datasets, in the same way the community has used Stable Diffusion and Flux models.

Baidu released Unlimited-OCR, an open-source 3B mixture-of-experts OCR model designed for long-horizon document parsing. This document model can parse 40 or more pages in a single forward pass by using a sliding attention window and constant-KV-cache to avoid memory and latency growth on long documents. This makes it practical self-hosted alternative for local OCR rather than sending documents to a large general-purpose API.

OpenAI made Codex Remote broadly available across ChatGPT plans, allowing users to initiate, continue and supervise coding work on connected Windows or Mac computers from the ChatGPT mobile application.

OpenAI expanded its Daybreak cybersecurity platform, introducing a full version of GPT-5.5-Cyber and new Codex Security capabilities designed to move from merely discovering vulnerabilities toward developing, testing and deploying patches.

OpenAI announced Codex Record and Replay, a workflow capture feature for desktop automation. Codex records clicks and actions, then generates an editable workflow file it can replay, which provides more adaptive software automation than relying on fixed pixel coordinates.

AI Research News

Google finds that reasoning can help language models retrieve facts they already know. Google researchers studied why chain-of-thought generation sometimes improves answers to straightforward factual questions that seemingly require no multi-step reasoning. Their analysis identified two mechanisms: reasoning tokens give the model additional opportunities for internal computation, while generating related information can prime the model to retrieve the correct fact, suggesting that reasoning improves memory access as well as formal problem-solving.
Source:

Google accelerated Gemini Nano with frozen multi-token prediction. Google Research described a technique that adds multi-token prediction to an existing on-device model without retraining the entire network. By freezing Gemini Nano’s main parameters and training lightweight prediction components to anticipate several tokens at once, the researchers increased generation speed on Pixel devices while avoiding the cost and complexity of rebuilding the underlying model.

AI Business and Policy

OpenAI and Broadcom announced the Jalapeño inference processor, a custom “Intelligence Processor” chip co-designed by Broadcom and OpenAI for LLM inference. Designed in nine months with AI assistance, the Jalapeño chip is designed specifically to support the serving patterns of LLM inference to deliver significantly better performance per watt than current systems and reduce OpenAI’s cost of serving models. This is an inference chip rather than a training chip, so OpenAI will still rely on Nvidia GPUs for training models.

Reflection AI signed a multibillion-dollar SpaceX compute agreement, where Reflection AI will pay SpaceX $150 million per month for access to Nvidia GB300 systems at the Colossus 2 data center. The contract could be worth as much as $6.3 billion through 2029.

Google DeepMind is losing key researchers to Anthropic. Gemini researchers Jonas Adler and Alexander Pritzel left Google for Anthropic, following the recent exits by Noam Shazeer and DeepMind director John Jumper.

Meta halted worker tracking for AI training due to privacy fears. Meta had started tracking workers’ computer usage to create AI training data just two months ago, but it created an employee backlash and led to data being accessible outside of the company.

Patronus AI announced a $50 million Series B funding round. The startup utilizes simulated digital environments to stress-test the reliability of AI agents performing complex, multi-step tasks.

General Intuition raised $320 million at a $2.3 billion valuation. The startup is developing agentic models and world models by leveraging action-labeled gameplay data from Medal to train for spatial-temporal reasoning in simulation and robotics.

Adobe acquired Topaz Labs to integrate Topaz AI models for video and image enhancement into its Firefly AI app and various creative suites.

OpenAI called for shared institutions to govern advanced models in a policy paper, OpenAI argued that governments need stronger technical institutions capable of evaluating frontier models, protecting sensitive systems and developing common standards for increasingly powerful AI. The proposal emphasizes evaluation capacity, international cooperation and repeatable release procedures as governments become more involved in reviewing models with advanced scientific or cybersecurity capabilities.

AI Opinions and Articles

The Atlantic launched a searchable public database called the AI Watchdog to track copyrighted material used to train AI music algorithms. The free tool allows musicians and content creators to search for specific artists or channels to see if their intellectual property was utilized by platforms like Suno or Udio. The database reveals that training datasets contain extensive quantities of content from prominent musical artists as well as hundreds of videos from popular independent YouTubers.

Top Tools

Z.ai launched GLM-5.2, a 753-billion parameter Mixture-of-Experts (MoE) open-weights AI model for long-horizon coding and engineering tasks. GLM-5.2 features a 1-million-token context window, reasoning controls, and support for coding tasks across entire codebases; it also utilizes the IndexShare architecture to reduce per-token compute FLOPs by up to 2.9 times. GLM-5.2 demonstrates high performance on benchmarks like SWE-bench Pro (62.1%), Terminal-Bench 2.1 (81.0%), and FrontierSWE (74.4%), rivaling frontier models like GPT-5.5 and Claude Opus 4.8.

Independent evaluations by Artificial Analysis confirm GLM-5.2 as the leading open weights AI model on the Artificial Analysis Intelligence Index. It also shows GLM-5.2 is notably token-hungry, consuming roughly 43,000 output tokens per standard Index task, up from 26,000 tokens used by GLM-5.1. GLM-5.2 is priced at only $1.40/$0.26/$4.40 per 1M input/cache hit/output tokens, so despite the token-hungry reasoning, this is the lowest-cost frontier-level AI model, substantially cheaper than proprietary rivals.

GLM-5.2 is positioned as a powerful open model focused on agentic software engineering that developers can run and build on. The GLM-5.2 model is available through Z.ai’s Coding Plan, their ZCode agent, their Z.ai chatbot, and via open weights on HuggingFace.

AI Tech and Product Releases

Sina Weibo researchers released VibeThinker-3B, a 3B parameter model matching flagship reasoning performance. The model achieves a score of 94.3 on AIME 2026 and 80.2 LiveCodeBench v6, a remarkable out-performance for a 3B model that matches top-tier AI models such as Claude Opus 4.5. This has stimulated conversation about benchmark and scaling limits.

The Technical Report on VibeThinker-3B shows that verifiable reasoning capabilities can be compressed into much smaller models than those required for open-domain knowledge, calling this the Parametric Compression-Coverage Hypothesis. This could have huge implications for how much further we can compress AI reasoning and improve AI model efficiency.

xAI released Grok Imagine Video 1.5 to general availability, xAI’s image-to-video system for generating short clips with synchronized sound. The update features improved motion physics, better audio synchronization (same-pass audio and speech generation), and nearly doubled generation speeds for 720p videos; its fast mode produces 6-second 720p videos in about 25 seconds. The release also adds workflow features such as Projects, multiple parallel agents, and search. Results are compelling:

You type a prompt or upload an image, and it turns it into a realistic 720p video, up to 15 seconds long, with actual dialogue and sound effects. All in just 25 seconds.

Google made Gemini Omni available through an API and positioned it as a leading video model. Gemini Omni is Google’s unified any-to-any system for text, image, video, audio, and music generation and editing. Google’s model page says Omni performs strongly on video editing, text-to-video, image-to-video, and reference-to-video, and reports top results on MovieGenBench for overall preference and instruction following. The model is meant for iterative multimodal video creation, including continuation, reference-based edits, and consistency across turns.

Anthropic has overhauled Claude Design, introducing enhanced canvas controls for easier element manipulation and brand-compliant design system imports from GitHub or local files. The update expands integration and export capabilities to platforms like Adobe, Canva, and Vercel, while also implementing shared usage limits across Anthropic’s product suite. Additionally, a new `/design-sync` command enables seamless, bidirectional workflow synchronization between Claude Design and Claude Code.

Nvidia released XR AI in public beta, a developer framework for building multimodal AI agents that run on AR glasses and extended-reality devices. The system connects video, audio, depth, pose, and sensor data with enterprise retrieval, AI models, agent orchestration, and accelerated inference. This enables hands-free AI assistance in laboratories, factories, hospitals, and design workflows.

HumanLayer launched its Agentic IDE for teams working in complex codebases. Declaring they are on a mission to ‘solve the AI slop code problem’, HumanLayer is aiming their HumanLayer Agentic IDE at structured, team-based AI software development rather than one-shot vibe coding. It includes a collaboration platform and software-factory building blocks designed to help engineers ship 3x faster while maintaining code quality and standards.

OpenRouter introduced Fusion, which combines access to multiple AI models behind one API call. Fusion sends one prompt to a panel of models, has a judge model compare the outputs, and then synthesizes a final answer from the combined results. In OpenRouter’s reported DRACO tests, fused panels outperformed individual models, and a lower-cost panel came within about 1 percentage point of Fable 5 while costing about half as much.

Midjourney announced Midjourney Medical, a new business to build a full-body ultrasound scanner called Ultrasonic CT that can do whole-body ultrasound scans in as little as 60 seconds, then offer it as a service in Midjourney Medical spas. The ultrasound system uses thousands of ultrasonic transducers to build a 3D anatomical map. It seems like a big leap to go from image generation into medical imaging hardware and services, but Midjourney Medical notes that large data volumes would make AI useful for processing and reconstruction.

Samsung announced a new AI-powered pet health feature for mobile devices during the VivaTech 2026 conference in Paris. Developed in collaboration with the platform Lifet, the feature uses AI to analyze photos of pets to detect conditions such as obesity and periodontal disease.

Tokyo-based AI startup Sakana AI has launched its first commercial product, Sakana Marlin, an autonomous AI research agent that works for up to eight hours to deliver deeply researched 100-page strategy reports and executive slides. The Sakana Marlin platform is designed exclusively for enterprise use and features a strict data policy ensuring customer inputs are never used for model training without consent.

OpenAI updated its platform deprecation notices for older GPT-5 and o3 model snapshots. Older GPT-5 and o3 snapshots will be removed from the API on December 11, 2026, while older GPT Image models and other legacy model families also have scheduled removals.

AI Research News

Researchers from multiple US Universities released SciAgentArena, a benchmark for evaluating AI agents in realistic scientific research scenarios. The benchmark includes roughly 200 tasks with stepwise verification, and an interactive agent-agnostic environment to assess AI agents. Benchmark results show that current agents are useful for well-specified data-analysis workflows but weaker at novel insight generation, exploration, and robust open-ended scientific reasoning.

Google DeepMind published “From AGI to ASI,” which explains Artificial Superintelligence (ASI) as systems surpassing large human organizations in capability and investigates the transition to improving AI from AGI to ASI. It explores four potential development pathways: scaling, paradigm shifts, recursive self-improvement, and multi-agent collectives. Each path has bottlenecks, and AI progress may accelerate continuously rather than in a single step change. Their roadmap shows ASI is attainable in the near future, requiring a global effort to prepare for coming transformative societal shifts.

AI Business and Policy

SpaceX bought Cursor’s parent company, Anysphere, in a $60 billion stock deal. The acquisition followed the massive SpaceX IPO and consolidates the AI landscape, bringing Cursor’s AI coding application and data into xAI’s broader AI model and AI infrastructure efforts. This acquisition integrates Cursor’s user base into SpaceX’s AI unit while bolstering Cursor’s position, as Cursor’s market share among AI coding tools slipped to 26% amid intense competition from tools like Claude Code.

Enterprise software ecosystems are undergoing an aggressive shift in pricing structures as CIOs push back against traditional seat-based subscription models in favor of consumption-based or outcome-focused metrics. Because autonomous AI agents operate independently of human headcount, software vendors are rewriting their commercial terms to charge based on API token volume, compute utilization, or verified task completion.

DeepSeek raised more than $7.4 billion in a funding round that valued the company at more than $50 billion, making it the most valuable Chinese AI startup. Its founder, Liang Wenfeng, invested around $3 billion in the fundraise. He previously held nearly 90% of the company before the financing round. A government-backed fund invested around $150 million.

At the G7 meeting, French President Emmanuel Macron urged the U.S. to share cutting-edge AI and called for democratic cooperation on regulation, in the wake of U.S. restrictions on Anthropic’s Fable 5 and Mythos 5 models. Macron criticized unilateral restrictions on Anthropic’s models as too nationalist.

Likewise, European Commission President Ursula von der Leyen said it is in both U.S. and EU interests for Europe to have access to the best AI models. The EU wants shared access to frontier AI capabilities under common safety standards rather than a drift toward nationalist AI controls.

OpenAI CEO Sam Altman and other AI leaders supported an international coalition for AI safety standards at an AI CEOs and leaders meeting at the G7 summit, with AI tech leaders proposing international cooperation with democratic oversight over AI deployments.

Anthropic and Tata Consultancy Services announced a partnership to bring Claude to regulated industries. TCS will provide Claude to 50,000 employees in 56 countries, build Claude-powered products for financial services, healthcare, public-sector, aviation, telecom, and life-sciences clients, and join the Claude Partner Network.

Anthropic also announced a multi-year global alliance with DXC Technology. DXC will train tens of thousands of Claude-certified forward-deployed engineers and integrate Claude into systems used by banks, airlines, insurers, manufacturers, and government agencies. DXC notes that Claude was used to generate more than 95% of the code for DXC OASIS, its AI-native managed-services orchestration platform.

AI Opinions and Articles

Jeff Bezos argued AI will ultimately create labor shortages rather than mass unemployment. Speaking at VivaTech, Bezos framed AI as a productivity accelerator that will expand the economy and create new kinds of work, a sharply optimistic contrast to surveys showing widespread job-loss concern. His argument captures executive optimism about how AI is a door to more opportunities rather than thinking of the economic possibilities as static.

Top Tools

Anthropic launched Claude Fable 5, its first public Mythos-class model, alongside the highly restricted Claude Mythos 5 model. Fable 5 is easily the most intelligent AI model yet released, with SOTA benchmarks on knowledge work (1932 on GDPval-AA), agentic coding (80.3% on SWE-Bench Pro, 88% on TerminalBench 2.1), reasoning (59% on Humanity’s Last Exam), and top positions across several capability leaderboards, including Agent Arena.

These models are excellent for long-horizon agentic work, use cases like Riley Brown re-implementing the whole Lovable interface in 2 prompts bear this out.

Positioned as a premier autonomous agentic system made safe for general use, Anthropic introduced significant safeguards on Fable 5, by redirecting high-risk cyber, biology, chemistry, and model-distillation requests away from the frontier model to Opus 4.8.They shared details in their Claude Fable 5 and Claude Mythos 5 System Card.

But Anthropic went further in the original Fable 5, silently sabotaging Fable 5 on requests that might relate to competitive AI development. This hidden output degradation led to significant backlash and criticism about trust and evaluation integrity. Anthropic then reversed course and changed the behavior so flagged requests visibly fall back to Opus 4.8 with explicit reasons for API users.

But for now, Fable 5 is gone. Anthropic abruptly suspended access to Claude Fable 5 and Claude Mythos 5 after it received a U.S. government export-control directive barring access by foreign nationals. Because the order applied broadly to foreigners, including foreign-national Anthropic employees, the company said it had to abruptly disable the models for all customers.

In his recent “Policy on the AI Exponential” blog post, Dario Amodei advocated for an Advanced AI Framework for overseeing models that would allow Government to block AI model releases. However, Anthropic insists that the Government’s ban on Fable 5 is based on a misunderstanding of its model’s risks due to a report of a jailbreak. They say, “We believe this is a misunderstanding and are working to restore access as soon as possible.”

AI Tech and Product Releases

Apple used WWDC26 to unveil their next generation of Apple Intelligence and a rebuilt Siri AI.

Apple announced its third generation of Apple Foundation Models, a family of five custom models developed in collaboration with Google. The AFM 3 models include on-device models and cloud models:

• AFM 3 Core, a 3B dense model for on-device use.

• AFM 3 Core Advanced, a multimodal 20B parameter sparse MoE for multimodal device tasks.

• AFM 3 Cloud, a server-side workhorse model for speed and performance.

• AFM 3 Cloud Image, for image generation and editing in photo-editing and Image Playground.

• AFM 3 Cloud Pro, for demanding agentic and reasoning use cases.

The AFM 3 framework is designed to power contextual, multi-platform AI experiences across the Apple ecosystem, leveraging both on-device hardware and Apple’s secure private cloud servers.

To support custom AI, Apple is also introducing Core AI, a new framework for running custom AI models on Apple silicon and Apple devices.

Apple presented the new Siri AI as far more capable at using personal context, app actions, and on-screen information, and more deeply integrated across Apple’s products - iPhone, iPad, Mac, Apple Watch, and Vision Pro. The new Siri AI adds web-based world knowledge and Visual Intelligence, and it is built around App Intents and App Schemas so apps can expose content and actions in natural language. This helps the new Siri AI handle multi-step requests like finding specific photos, organizing emails, and taking actions across apps.

Apple updated Image Playground with their newest AI model, AFM 3 Cloud Image, so it is capable of producing improved photorealistic and stylized graphics, more in line with competitive image generation tools.

Google launched DiffusionGemma, an experimental 26B Mixture of Experts (MoE) model with 3.8B active parameters that uses text diffusion and is released under an open-source Apache 2.0 license. DiffusionGemma utilizes text diffusion to generate 256-token text blocks simultaneously, which yields up to six times faster local generation speeds (over 1,000 tokens per second on a single H100). The speed makes it ideal for interactive workflows like in-line editing of code and documents, but it has lower overall output quality compared to the 26B Gemma 4 model.

Google introduced Gemini 3.5 Live Translate, a speech-to-speech model for near real-time voice translation in more than 70 languages. Gemini 3.5 Live Translate features a single continuously streaming audio model rather than a stitched pipeline of speech recognition, translation, and text-to-speech components. This model preserves tone, pacing, and expressiveness while translating with sub-second latency. This update is rolling out to Google Translate and Google Meet for live translation, with developer integrations available via the Gemini Live API.

Moonshot AI released Kimi K2.7-Code, an open-source update on the Kimi K2 1T parameter MoE architecture that claims a 30% reduction in thinking-token usage compared to its K2.6 predecessor. K2.7-Code is an open weights model available on HuggingFace and available via the Kimi Code platform. While Moonshot AI reports performance gains on internal benchmarks, independent evaluations on KernelBench-Hard showed regressions in specific GPU kernel optimization tasks. Elliot Arledge’s benchmarking assessment is “K2.7 is more honest but not more capable” than its K2.6 predecessor on Cuda kernel coding.

Cohere released North Mini Code, an open-source 30B agentic coding model aimed at developers who want AI coding agents that can be run and improved outside closed proprietary systems. The model is the company’s first model for developers and is available on HuggingFace under the Apache 2.0 license.

Google introduced new Gemini features tailored for small businesses, including a direct Google Business Profile connection and proactive Business notebooks. The update allows Gemini to integrate with Google Business Profiles to access customer reviews, questions, and performance data. New Business notebooks provide a centralized space to organize workflows and generate content based on specific business context.

Cognition launched FrontierCode, a tougher coding benchmark designed to measure whether an AI-generated pull request is production-quality. Built from 150 original tasks, the benchmark emphasizes evaluation criteria such as scope control, regression safety, and test quality. Fable 5 posted the highest score (46.3%) versus Opus 4.8 (34.3%) and GPT-5.5 (25.5%).

Xiaomi’s MiMo AI team has open-sourced MiMo Code V0.1.0, a terminal-native AI coding harness that Xiaomi claims can outperform Claude Code on long tasks. The assistant utilizes a cross-session memory architecture with a dedicated checkpoint-writer subagent to maintain context during long-horizon, multi-step tasks.

Avataar AI from India launched a new video model called Varya that uses distillation from Alibaba’s Wan 2.2 and generates video ten times faster than the original at a cost of under a penny per second. Varya features Indian cultural nuances By tuning the model with curated data for the India market. It can be accessed at the Varya platform and will be released as an open-weight model on the India’s AIKosh portal.

Deezer introduced a tool to identify AI-generated tracks in streaming playlists. The free online detector supports 27 languages and scans music from 20 platforms, including Spotify, Apple Music, and YouTube Music. Deezer reports that 44% of all new music uploaded to its platform is AI-generated.

AI Research News

Google recently published “Accelerating scientific discovery with Co-Scientist” in Nature, an account of how the Co-Scientist AI system is designed to help solve complex problems in the life sciences. The tool uses specialized agents to generate, debate, and refine new hypotheses through three distinct phases of idea generation, peer review, and refinement. One case of Co-Scientist was helping to identify new drug repurpose candidates and synergistic combination therapies for acute myeloid leukemia.

UC Berkeley researchers launched Agents’ Last Exam (ALE) Benchmark, which evaluates AI agents on long-horizon professional workflows. OpenAI’s GPT-5.5 leads the ALE Leaderboard with a 24.0%, beating Anthropic’s Claude Fable 5.

Artificial Analysis developed AA-AgentPerf, a new hardware performance benchmark that measures how many concurrent agentic AI agents a system can sustain. Nvidia’s GB300 sets a new standard for agentic AI workload performance, over 20 times better than the H200.

AI Business and Policy

SpaceX launched their IPO into the stock market stratosphere, rising on the IPO debut to a valuation over $2 trillion by the market close. SpaceX is rising in part due to XAI, its Colossus AI data center it is renting to Anthropic, and their claims of pursuing a $4 trillion AI opportunity.

SpaceX is combining their AI and space opportunity with a proposed satellite designed to host AI supercomputers in orbit. The engineering specs describe about 150 kW peak power per satellite, and Musk claims the challenge it not harder than some other things they are doing. I doubt this idea will come to fruition soon; it seems it has been hyped up lately for IPO buzz.

OpenAI announced that it had confidentially submitted a draft S-1 to the SEC, giving the company the option to go public while emphasizing that timing has not been decided. The announcement said OpenAI expects the filing to leak and still sees tradeoffs between remaining private and preparing for a public offering.

OpenAI announced that it will acquire Ona, a company focused on secure cloud execution and orchestration. Ona’s technology will help Codex expand from a session-bound developer tool into a persistent agent environment that supports long-running software and knowledge-work tasks.

OpenAI and Oracle announced that OCI customers will be able to access OpenAI models and Codex using existing Oracle cloud commitments.

OpenAI announced support for the EU Code of Practice on Transparency of AI-generated content, tying the move to its provenance and content-authenticity work. This guides AI providers in standards and tools to help users distinguish synthetic media from human-created content.

Apple said that Siri AI will be delayed on iOS 27 and iPadOS 27 in the European Union because of the Digital Markets Act. Apple said EU users will still be able to access Siri AI on macOS 27 and visionOS 27, but that iPhone, iPad, and watchOS access will not arrive on the same timeline because of unresolved regulatory concerns.

OpenAI published a threat report claiming PRC-linked influence operations are targeting AI debates in the United States, including around data center buildout. The OpenAI threat report frames the incidents as Chinese-based covert influence operations aimed at shaping political and public opinion. The report relates this to a spike in anti-datacenter social media activity and says OpenAI identified accounts using ChatGPT as part of those campaigns.

Anthropic introduced Claude Corps, a national fellowship program for early-career people interested in using AI for public benefit and community impact.

AI Opinions and Articles

Reuters reports on the broad public anxiety about AI among the US public, citing a Reuters/Ipsos poll showing high levels of concern about AI use and job displacement. With AI usage skyrocketing and OpenAI and Anthropic moving toward public listings, the investor appetite for AI companies is colliding with greater public unease about AI’s economic effects.

If you have AI fears or AI FOMO, the best move is to learn AI. OpenAI introduced three new OpenAI Academy courses: AI Foundations, Applied AI Foundations, and Agents and Workflows. The courses are available to ChatGPT users and will help individuals and organizations move to using AI in repeatable AI workflows and agent-assisted work.

Top Tools

The top AI model announcement for this week has been Microsoft announcing a new family of seven in-house MAI models at Build 2026. The new AI model lineup spans reasoning, coding, image generation, transcription, and voice, and includes MAI-Thinking-1, MAI-Code-1-Flash, MAI-Image-2.5, MAI-Transcribe-1.5, MAI-Voice-2, and Flash variants for image and voice.

Microsoft launched MAI-Thinking-1 is a Mixture of Experts model with 1T total parameters and 35B active parameters; it’s their flagship reasoning model. Comparing it to Sonnet 4.6, Microsoft says it was trained from the ground up without third-party distillation and is competitive in its class on coding (52.8% on SWE-Bench pro, but only 46% on Terminal Bench 2.0) and mathematical reasoning benchmarks.

Almost as impressive as the model itself is Microsoft’s 109 page technical report “MAI-Thinking-1: Building a Hill-Climbing Machine,” which shares details on the model architecture and how Microsoft AI trained their model. This is the most open an American AI lab has been about their work in some time.

Microsoft also highlighted MAI-Image-2.5, including a Flash variant, as its new image model that ranks number two on Arena for image editing. Microsoft is rolling it out to support PowerPoint visuals and OneDrive Photos editing tools.

Some of Microsoft’s other Build announcements:

• Microsoft unveiled the Surface RTX Spark Dev Box that uses Nvidia’s RTX Spark superchip to enable users to run powerful AI models on local Windows machines.

• Microsoft introduced their ‘always on’ AI agent Microsoft Scout, their entry into the general local AI agent, based on the OpenClaw framework and OpenShell.

• Microsoft introduced Work IQ APIs, a context layer for autonomous task execution and enterprise customization that leverages Microsoft’s ecosystem with Work, Fabric, Foundry, and Web IQ components.

• Microsoft unveiled the Majorana 2 quantum chip that features qubits that are 1,000 times more reliable than previous generations.

• Microsoft introduces Microsoft Execution Containers (MXC) to secure AI agents. MXC is a policy-driven execution layer built into Windows that allows developers and administrators to define sandbox environments and enforce access boundaries for AI agents.

The bigger picture is that Microsoft is making a strategic shift toward in-house superintelligence development, directly challenging leading AI labs by becoming one. Microsoft also found itself behind the curve with its chatbot-based Copilot suite and is now trying to catch up with AI agent offerings and support.

AI Tech and Product Releases

Nvidia made several announcements at Computex, including several new AI models. Nvidia released Nemotron 3 Ultra, a 550B parameter sparse MoE open-weights model with 55B active parameters, designed for long-context and agentic workloads. This Mixture-of-Experts model utilizes a hybrid Transformer-Mamba architecture, which supports a longer 1 million tokens of context as well as faster and lower cost inference for agentic workloads. Nemotron 3 Ultra and is being released with model weights, training assets, datasets, and related tooling. It is available on Amazon SageMaker JumpStart and other platforms.

Nvidia also released Nemotron 3.5 ASR, a 600M parameter multilingual streaming speech recognition model. This model uses a cache-aware FastConformer-RNNT architecture to deliver high-quality speech-to-text in both streaming and batch mode transcriptions. Nemotron 3.5 ASR supports 40 language locales and adds punctuation and capitalization to transcripts.

Nvidia announced RTX Spark, a new Windows PC platform built for local AI agents. Built on the same hardware used in Nvidia’s DGX Spark, the Windows RTX Spark delivers up to 1 petaflop of AI performance and up to 128GB of unified memory, offering the ability to run large and powerful AI models on Windows laptops and compact desktops.

We wrote more on RTX Spark in “RTX Spark: AI Comes Home to the PC” as well as other Nvidia announcements, including the release of Cosmos 3, the latest iteration of their open-source frontier omni model for physical AI, which integrates world generation, physical reasoning, and action generation into a single framework.

MiniMax announced M3, a native multimodal AI model with a 1 million token context window that is frontier-class at coding and agentic AI. Minimax touts M3’s impressive benchmarks such as 59.0% on SWE-Bench Pro and 66.0% on Terminal Bench 2.1, competing with Gemini 3.1 Pro and GPT 5.5. Minimax promises a fuller technical release with open weights in 10 days. When it does release, it will be SOTA for open weight AI models. In the meantime, access is via their API ( at $0.60 / $2.40 per million input / output ) and on the Minimax platform.

OpenAI upgraded GPT-Rosalind for life sciences with new GPT-Rosalind capabilities aimed at enterprise-scale biology, genomics, medicinal chemistry, and drug-discovery workflows. The new GPT-Rosalind model release pairs GPT-5.5-style coding and tool use with life-sciences reasoning, adds research and analysis plugins in Codex. This places GPT-Rosalind as a domain-specific scientific AI workbench with provenance, tools, and controlled access. OpenAI is expanding access to eligible research organizations through a trusted-access model.

OpenAI is significantly updating its Codex agentic AI platform for non-coding knowledge work. Codex is adding six role-specific plugins - data analytics, creative production, sales, product design, public-equity investing, and investment banking - that integrate over 60 business applications, such as Salesforce and Figma, to automate complex enterprise workflows.

Codex is also being updated with a Sites features for hosting interactive, semi-private web applications and an Annotations feature for in-place content editing and refinement. These enhancements are designed to expand Codex’s utility for non-technical workplace users by deeply integrating it with existing professional tools and workflows. Codex has more than 5 million weekly users and non-developers make up about 20% of usage.

Google released Gemma 4 12B, a new natively open multimodal model in the Gemma 4 family. With 12B parameters, performance that is SOTA for its size, and ability to process audio natively, it is a great laptop-runnable AI model for local AI use. Google also updated their Gemma 4 lineup with Quantization-Aware Training (QAT) to reduce memory footprints and help quantized Gemma models perform better.

Ideogram released Ideogram 4.0 as its first open-weight text-to-image diffusion transformer foundation model. Ideogram 4.0 he model is a 9.3B-parameter text-to-image system trained from scratch, uses Qwen3-VL-8B-Instruct as its text encoder, and is built around structured JSON prompts with optional layout and color controls.

Reve released Reve 2, a new text-to-image model centered on layout-aware generation and editing. As Reve says:

Layout is a structured, hierarchical description of an image where every element has a location, a size, a local description, and other optional attributes like image references or color. A layout is an image’s backbone — separating semantic intent from pixel rendering, much like HTML is to a webpage or SVG to a vector image.

Reve says the system separates planning from rendering, represents images in a structured form that makes individual elements addressable, and renders at native 4K resolution for more precise editing control.

LMArena launched Agent Arena, a benchmark and comparison platform for AI agents rather than one-shot chat prompts. The platform evaluates agents built from models, tools, and frameworks across real-world tasks, and the launch included a public release of 2,000 pairwise agent battles and user preference data.

ChatGPT is rolling out a new memory architecture called Dreaming, a more scalable memory synthesis system for ChatGPT designed to keep user context fresh, relevant, and correct over longer time periods. OpenAI says Dreaming improves how ChatGPT carries forward context, follows user preferences, and stays current. Memory for AI is evolving from explicit saved notes toward automated synthesis across conversations. The feature is available to Plus and Pro users in the U.S., with broader rollout planned for coming weeks.

JetBrains open-sourced Mellum2, a 12B MoE model with 2.5B active parameters per token that is positioned for production AI workloads such as routing, summarization, and intermediate reasoning over natural language and code. The model was built from scratch and released under Apache 2.0 with weights available on HuggingFace.

H Company launched Holo3.1, a family of local computer-use agent models ranging from 0.8B to 35B-A3B. H Company says the release improves robustness across web, desktop, and mobile environments, and raises AndroidWorld results from 67% to 79.3% on its 35B-A3B model.

Alibaba released Qwen3.7-Plus, a multimodal model with frontier-level performance and a 1-million token context window. It is 60% cheaper than the previous text-only Qwen3.7-Max, but the release marks a departure from Alibaba’s open-source strategy, as the model is available only via proprietary APIs.

AI Research News

Anthropic published “When AI builds itself,” a report on recursive self-improvement and AI-driven AI R&D. AI has progressed from chatbot to single agents to multiple autonomous AI agents, and now Anthropic lays out the next step, where AI agents “close the loop” and AI development becomes substantially automated. They show early evidence of this trend by noting Anthropic itself is shipping 8 times more code per person than in previous years. This near-future AI trend implies both AI acceleration and huge leaps in productivity in some companies.

AI Business and Policy

Anthropic has filed confidentially for an IPO following an oversubscribed $65 billion fundraise at a $965 billion valuation. The IPO depends on SEC review and market conditions, but it is expected later this year and could be the second trillion-dollar IPO following the SpaceX IPO. Anthropic’s annualized revenue surpassed $47 billion in May, up from roughly $9 billion at the end of 2025.

SpaceX is becoming a hyperscale AI compute provider as it preps for its historic IPO on June 12. SpaceX has secured a deal with Google to provide approximately 110,000 Nvidia GPUs and related components from October 2026 through June 2029. Google will pay $920 million per month to secure bridge capacity for surging demand on its Gemini Enterprise AI platform, follows a similar computing agreement between Anthropic and SpaceX.

OpenAI called for global action on youth AI safety through a dedicated AI Safety Institute. OpenAI is advocating for the establishment of an international institute to provide continuous oversight and standardized guidance for youth AI safety.

President Trump signed an executive order on AI innovation and AI security. The new Executive Order NSPM-11 emphasizes promoting appropriate AI adoption and AI innovation for national security, while coordinating with the private sector on security risks. It also directs federal agencies to prioritize AI-related cybersecurity, establish an AI cybersecurity clearinghouse with voluntary industry collaboration, and expand federal cybersecurity hiring pathways.

Senior U.S. officials have been discussing with AI firms such as Open AI the potential for the federal government to acquire shares in those companies. Giving the U.S. government equity stakes in AI companies could have seismic consequences. While the arrangement could fund public purposes like dividend payments, critics warn that government ownership could create conflicts of interest in technology regulation.

OpenAI and other AI leaders have backed synthetic DNA screening rules, with executives and scientists from leading US AI labs signing a letter urging U.S. lawmakers to require screening of synthetic DNA and RNA orders. The DNA-screening letter expresses the concern that advanced AI is lowering the knowledge barrier for designing dangerous biological materials, making gene-synthesis oversight more urgent.

Character.AI and Google reached settlements with families over teen suicide claims, notifying a Florida federal court of a mediated settlement to resolve all claims. The litigation included a high-profile lawsuit alleging that the chatbot encouraged a 14-year-old to commit suicide.

Meanwhile, OpenAI is responding to a lawsuit filed by the family of a teenager who died by suicide, arguing that the chat logs used in the allegations ‘require more context.’ They also claim the chatbot frequently directed the user to crisis resources and that the incident resulted from improper use of the platform.

The Linux Foundation unveiled plans for the Tokenomics Foundation to address rising AI token costs. The Tokenomics Foundation charter is to establish open industry standards, benchmarks, and best practices for the economical use of AI infrastructure. The new standards body aims to establish a framework for tracking, auditing, and optimizing AI token usage and billing.

The New York State legislature passed a one-year moratorium on large new data centers. The bill directs an environmental agency to assess the electricity, water, and land usage of large-scale facilities.

Meta has built data centers in tents to accelerate AI infrastructure deployment. The company has constructed six “rapid deployment structures” in New Albany, Ohio, to reduce construction time by half. These structures will house AI chips and are powered by 200 megawatts of modular gas turbines.

Jensen’s Computex Keynote

Nvidia CEO Jensen Huang gave another gangbuster keynote at Computex conference in Taiwan on Monday, declaring “useful AI has arrived” and calling this the Age of Agents, a new computing era where AI is delivered through running AI agents that reason, plan, and use tools.

Jensen Huang had real announcements behind his hype, introducing several new AI models, AI chips, and systems. Some of the key highlights:

• Nvidia announced the Vera CPU, explicitly designed for AI agent use and able to manage massive graphical processing units (GPUs) inside localized agentic loops. Vera integrates 88 custom “Olympus” ARM-based cores and utilizes LPDDR5X memory to achieve 40% lower peak memory latency, 50% faster core-to-core communication, and 1.8x performance over prior CPUs.

• Nvidia announced that the Vera Rubin system is now in full production. Vera Rubin is Nvidia’s latest generation of AI supercomputer that combines GPUs and CPUs specifically engineered for agentic AI. It marks a major milestone for Nvidia as it dominates as the leading AI infrastructure company.

• To support AI infrastructure buildout, Nvidia introduced the DSX blueprint, a reference design for building and operating AI factories that integrates hardware, cooling, power, and networking for maximum revenue-generating compute.

• Nvidia introduced Nemotron 3 Ultra, an open-source Mixture of Experts (MoE) AI model featuring 550B total parameters and 55B active parameters designed for agentic use. Nemotron 3 Ultra utilizes a novel hybrid State Space Model (SSM) architecture that makes is faster and cheaper for AI inference, yet with performance comparable to leading open AI models like MiniMax M2.7.

• Nvidia is providing an enterprise toolkit that includes models, orchestration harnesses like Open Shell, and access to CUDA X libraries that act as specialized skills for agents, to empower AI users to build their own agent harnesses and systems.

• Nvidia continues to push AI models and tools for physical AI, including updates to the Isaac Groot platform and a reference design for humanoid robotics. For the robotics, autonomous vehicle, and physical system sectors, Nvidia launched Cosmos 3, an “Omni” multimodal foundation model trained on 20 trillion tokens of images, audio, video, action data, and text. They also announced the Alpamo 2 open model designed for reasoning in autonomous vehicles, which enables level 4 autonomous driving for robotaxis.

However, the most interesting announcement was the RTX Spark Superchip and the Agentic PC, Nvidia’s effort to reinvent the personal computer with a 1 petaflop superchip that that Blackwell GPU and Grace CPU to run local AI agents natively.

RTX Spark Details

Fabricated on TSMC’s 3-nanometer process and packing 70 billion transistors, the system-on-chip RTX Spark merges a Blackwell-architecture RTX GPU featuring 6,144 CUDA cores with a custom 20-core Grace CPU. Boasting one petaflop of local AI performance and 128 GB of high-bandwidth unified memory, the hardware allows personal AI assistants to run locally, securely, and continuously inside Windows agent sandboxes.

Designed in collaboration with Microsoft and MediaTek, RTX Spark is engineered to support agent-centric computing on the consumer PCs and laptops and has integrated graphics and GPU performance comparable to a dedicated desktop RTX 5070 graphics card. Achieving that tier of graphics compute on an integrated platform represents a massive step forward for thin-and-light Windows laptops.

Consumer laptops and PC can be equipped with 32 GB of RAM for general productivity and gaming, but with top-tier configurations with 128 GB of unified memory you can run larger AI models and agentic AI systems natively.

With its unified memory design and integrated GPU, Nvidia’s RTX Spark gives the Windows ecosystem a competitive performance and efficiency equivalent to Apple’s series of M series of chips. Finally, there will be an AI PC that can give Apple M5-equipped MacBook Pro a run for its money.

The AI PC, Again

This isn’t the first time an ARM-based Windows PC has been attempted. The first ARM Windows device was Surface RT in 2012; later came Windows 11 on ARM powered by Qualcomm’s Snapdragon chips.

Nor is it the first AI PC. After the ChatGPT moment, Intel launched their bid for the AI PC chipset while Microsoft offered the Copilot + PC. The problem was these prior efforts lacked the performance needed for rigorous AI workloads and were eclipsed by better performance out of Apple’s MacBooks.

Nor is it Nvidia’s first attempt at an AI desktop box. Nvidia has been shipping the DGX Spark since 2025, a Linux-based desktop box for AI that uses a GB10 Grace-Blackwell chip with similar specs to the RTX Spark.

RTX Spark brings the Windows OS that DGX Spark lacks, turning NVidia’s Grace-Blackwell superchip into an engine for Windows laptops and PCs. This RTX Spark is a significant performance boost from prior ARM PC efforts, which gives it a better chance to crack that market.

One thing in favor of RTX Spark is that it’s built for gaming and graphics applications as well as AI workloads. Leveraging Nvidia Blackwell-class GPUs and its Cuda stack, the RTX Spark offers Deep Learning Super Sampling (DLSS) upscaling and frame generation technologies to accelerate games. Microsoft has worked with Windows tools developers to make their apps work on ARM-base Windows machines. Nvidia has also pushed Adobe to rebuild its Adobe Premiere Pro editing engine around GPU-accelerated computing rather than relying heavily on the CPU.

There are still many questions about it. Reviewers have noted that Nvidia did not share full benchmarks, leaving users to assess gaming performance on demo applications. Microsoft has announced Surface Laptop Ultra but it and other PC makers won’t be delivering them until later this year.

The main barrier to adoption could be cost. DGX Spark with GB10 chip and 120GB memory already costs $4700, and a similarly-loaded RTX Spark will undoubtably be in the same range, limiting widespread adoption. On the other hand, cutting the memory down to avoid high component pricing will limit the system’s utility with AI models.

Conclusion

With a market cap of about $5.4 trillion, Nvidia is worth more than any company on the planet, almost $1 trillion above its closest U.S. peers. It has placed itself at the center of the AI revolution by making critical and correct bets about the technology stack needed to support AI.

Nvidia has earned huge success building the core GPU chips and the AI supercomputers that go into AI factories in data centers. Now, Nvidia is going after the one led by others, the PC market, This bringing AI home to the PC. Wall Street is recognizing the threat it poses to others; this announcement at Computex on Monday moved Nvidia up and Intel and AMD down.

Still, behind the hype, what RTX Spark is really doing is bring ARM-based Grace-Blackwell CPU-GPU integrated SoC to Windows. Is this what users need? I think so, but buyers will decide and devices won’t be out until later this year when they’ll compete with Apple’s M6. It’s not a sure thing.

One Reddit commenter opined:

It [RTX Spark] is a pipe cleaner product - attempt to get this “NV-made Arm-based SoC for laptops” to market and be a thing for software developers. It is late, underpowered and probably overpriced (based on DGX Spark pricing and current RAM & storage prices) but it lays the groundwork for the next gen chip that might be better.

Whatever the market reaction, Nvidia will keep pursuing it, because Jensen’s vision of an “AI supercomputer for your home AI agent” is on target, and Nvidia will keep iterating until they get it right.

If and when it takes hold, we will likely see the end of the desktop PC graphics card and the rise of unitary memory, so larger AI models can be run on laptops and PCs. We will also see a class of home devices used not as personal computers but as Agent Computers, running OpenClaw or Hermes 24/7.

Nvidia moving beyond the data center and moving to the edge tells us something important: AI is reinventing the PC just like it is reinventing every part of the information technology stack. Smaller devices like phones and laptops will run advanced AI models and AI agents, literally bringing AI home.

Top Tools

Anthropic released Claude Opus 4.8, a frontier AI model upgrade to Opus 4.7, with stronger coding, agentic, and professional work performance. On benchmarks, Opus 4.8 achieves state-of-the-art 1890 on GDPval-AA for knowledge work and 69.2% on SWE-Bench Pro. Anthropic also touts improvements in its alignment and honesty, being less likely to hallucinate success or unverified claims.

The improvements over Opus 4.7 are solid but incremental, and they have kept standard pricing unchanged from the prior version. They also launched effort controls and a faster and cheaper fast mode that can be used for high-throughput workloads.

Anthropic also launched dynamic workflows in Claude Code for large tasks such as codebase-scale migrations. When users prompt for a complex task, Claude breaks the target down into subtasks and assigns sub-agents to the work. Claude dynamically runs tens to hundreds of parallel sub-agents in a single session, checking its work via internal agent critique before final output.

AI Tech and Product Releases

OpenAI launched Rosalind Biodefense, which gives trusted developers sponsored access to GPT-Rosalind for defensive biology work, including epidemiological modeling, early detection, screening, preparedness, diagnostics, and medical-countermeasure development. OpenAI is also expanding trusted access to selected U.S. government and allied public-health and biodefense partners.

Mistral introduced Search Toolkit in public preview. The open-source framework unifies ingestion, retrieval, and evaluation for production search pipelines used in AI applications. Mistral’s pitch is that teams should spend less time wiring together search infrastructure and more time improving retrieval quality; the toolkit can run in cloud, on-premises, or edge environments.

Mistral launched Vibe as Mistral’s live agent product and main AI interface, available through Mistral’s chat interface and mobile apps. Vibe now replaces LeChat and is absorbing prior Le Chat history, plans, and settings inside Chat mode. Vibe has a Work Mode AI agent for complex, multi-stage tasks, and a Code Mode as the new coding surface in the Vibe web app. The launch positions Mistral’s consumer and developer-facing agent around everyday tasks and knowledge work.

Mistral announced “physics AI” for industrial engineering. The company says it has brought Emmi AI into Mistral and is building AI models that learn from physics-solver outputs to predict physical fields from geometry, boundary conditions, or measurement data. The intended use cases include faster design-space exploration, tooling and process optimization. They aim to apply these physics AI models as real-time digital twins for industrial partners such as ASML, Airbus, Safran, and Siemens Energy.

Microsoft announced its new MAI Image 2.5 image generation model, an upgraded text-to-image generator succeeding MAI Image 2.0 that follows prompt instructions more closely and renders text strings more reliably. Climbing to the number three spot on the text-to-image Arena.ai leaderboard, MAI Image 2.5 displays strong visual reasoning around scenes and lighting, which combined with its sharper accurate text rendering makes it well-suited for branding and product concepts.

Microsoft rolled out an overhauled design for Microsoft 365 Copilot across its office productivity suite, calling it “a cohesive, agentic experience.” The new Copilot has a consistent entry point across apps and can now draw live data directly from other integrated Microsoft apps, such as emails, calendars, and files, to generate context-aware charts and graphs.

Microsoft is attempting to keep Copilot competitive as quickly evolving AI applications take on agentic abilities. To that end, Microsoft is reportedly developing a unified “super app” to consolidate GitHub Copilot, Copilot chat, and Copilot Cowork into a single destination. This new platform will feature an agentic workflow capability internally named Autopilot and is expected to launch by the end of summer.

Perplexity announced that its Perplexity Computer capabilities are now directly available within Microsoft 365 applications, including Word, Excel, and PowerPoint. The deep integration allows users to request multi-step, complex analytical actions beyond standard chat responses. For instance, the tool can analyze a legal document against a template, track changes, and generate an issues list with fallback clauses.

Eleven Labs released its upgraded Music V2 generative audio model, which focuses on producing higher-fidelity musical tracks. Eleven Labs claims:

Music v2 delivers better vocals, instrumentation, and arrangement across every genre, with improved multilingual support and a set of new capabilities.

The foundation model was trained entirely on licensed data, ensuring that commercial usage rights are cleared for content creators. Testing shows that the model contains built-in world knowledge, allowing it to correctly reference specific landmarks and pop culture elements when given regional prompts.

Eleven Labs also launched Dubbing V2, an automated video localization tool that translates audio content while preserving original attributes. The software takes an uploaded video file and converts the speech into one of over 90 target languages, translating while maintaining the speaker’s original vocal tone, emotional delivery, and facial expressions. This keeps the output more faithful to the original delivery.

Figma transformed its AI design assistant, Figma Make, into a live, visual software editor that connects natively to production codebases. The update allows users to import existing Git repositories directly into the Figma desktop app to visually edit underlying code and push changes back to engineering through GitHub pull requests. The platform utilizes a multi-model AI system, toggling between Anthropic’s Claude and Google’s Gemini models to write code that adheres to established design system guidelines.

MiniMax released a technical report on their M2 series and teased upcoming M3 models. The upcoming M3 series will feature “MiniMax Sparse Attention” (MSA), a sub-quadratic framework capable of 15.6 times faster decoding speed at million-token context lengths. The MiniMax-M2 Series Technical Report highlights the sparse Mixture-of-Experts architecture M2 and its training: Agent-driven data pipelines; the “Forge” reinforcement learning system for agent-native training; M2.7 taking steps toward self-evolution by autonomously debugging training runs.

Meta is developing an AI-powered pendant that it plans to start testing in the next year. The device is expected to build on the technology of Limitless, an AI startup acquired by Meta at the end of 2025. Meta also plans to expand its AI glasses lineup and launch a “Wearables for Work” business subscription.

OpenAI has added Codex’s computer use feature to Windows. The app can see your screen and perform tasks on your device. Users can also manage and review Codex’s jobs via the ChatGPT app.

OpenAI will remove Canvas feature in GPT-5.5 models. The side-by-side editing feature will no longer be available with GPT-5.5 Instant or GPT-5.5 Thinking. OpenAI is also shortening GPT-5.5 Instant responses and reducing the use of bullets in text.

AI Research News

The paper “Reasoning-preserved Efficient Distillation of Large Language Models via Activation-aware Initialization” argues that some efficient distillation methods damage multi-step reasoning through “reasoning collapse.” To fix this, the proposed RED method uses activation-aware initialization to better preserve hidden-representation rank. Experiments on Llama and Qwen models show that RED recovers reasoning while keeping the efficiency benefits of compressed LLMs.

AI Business and Policy

Anthropic raised $65 billion in Series H funding at a mind-boggling $965 billion post-money valuation, with Anthropic saying proceeds will support safety and interpretability research, compute expansion, and product scaling. Anthropic’s run-rate revenue crossed $47 billion earlier in May, leading OpenAI in revenue, and it has signed major compute agreements with Amazon, Google, and SpaceX to ramp up capacity for serving AI. Anthropic also opened a Milan office and expanded its European footprint.

OpenAI published its Frontier Governance Framework this week, which explains how OpenAI’s safety and security practices align with existing and emerging legal requirements, including in the US, California, and EU. The Frontier Governance Framework covers how OpenAI deals with AI risk assessment and mitigation in areas such as cyber offense, CBRN, harmful manipulation, and loss of control, providing guidance on model reporting, security management, and incident response.

The Verge examined the rapid normalization of AI in warfare in a feature that argues that military AI is no longer a future scenario. The article covered the shift from Project Maven to modern AI-enabled surveillance, object detection, and targeting workflows. tensions between government demand for broad “lawful use” and AI companies’ attempts to define ethical red lines around autonomous weapons and surveillance.

AI Opinions and Articles

Pope Leo XIV issued an encyclical letter on AI called “Magnifica Humitas”, which means ‘Magnificent Humanity’, with a focus on “safeguarding the human person” in the AI era. It’s a nuanced, informed, and detailed document covering the impact of AI and how we should approach it. The Pope emphasizes that humans possess a unique, inherent dignity that should not be overlooked as AI capabilities grow.

The Pope neither rejects AI in toto nor accepts the accelerationist argument but raises serious concerns and social impacts resulting from AI development, such as AI companionship’s impact on human relationships. He critiques how AI development being controlled by a few private entities complicates governing these technologies for the “common good.” The Pope advocates for “disarming” AI, meaning we must move away from a mentality of “armed competition” of the AI race and instead foster open, human-friendly collaboration.

Pope Leo XIV and the New Social Question of AI reviews Pope Leo XIV’s AI missive in the context of Pope Leo XIII’s Revum Novarum, which confronted challenges of industrialization over a century ago.

The Guardian scrutinized Anthropic’s association with Pope Leo XIV’s AI encyclical, sharing criticism that Anthropic’s engagement with the Vatican could become “Vatican-washing” if it burnishes the company’s safety image without addressing AI concerns. Anthropic is also using AI ‘concerns’ as a way to lock down AI development via ‘regulatory capture.’

The Pope has moved the AI ethics debate forward, addressing AI in religious, social, labor, and geopolitical contexts.

Top Tools

Google presented many AI updates at Google I/O this week, and we shared our breakdown of highlighted announcements and releases in a prior article. There were many AI announcements at Google I/O, but to recap our recap, these were the most important ones:

• Google introduced Gemini Omni, a multimodal world generation model that can create “anything from any input,” with natural-language editing across text, image, and video prompts, and video generation output.

• Google released Gemini 3.5 Flash and previewed Gemini 3.5 Pro. Gemini 3.5 Flash is positioned for agentic workflows, coding, long-horizon tasks, multimodal understanding, and real-time. Gemini 3.5 Pro is expected to roll out next month.

• Google introduced Antigravity 2.0, an agent-first platform that revamps Antigravity, and also unveiled the Gemini Spark personal agent, a 24/7 personal AI agent built on Gemini 3.5 and Antigravity.

• Google announced many AI-infused features across the Google ecosystem, including major AI updates for Search, personalized Daily Briefs, Universal Cart for AI-assisted shopping, Ask YouTube for video search, Google Pics for image editing, and intelligent eyewear powered by Gemini.

One way to summarize Google’s direction: Google expanded its agentic product layer across Search, Gemini, Workspace, shopping, YouTube, and Android XR. The Verge summarized Google I/O as a broad AI platform push across models, agents, apps, and hardware.

AI Tech and Product Releases

Alibaba’s Qwen Team released Qwen3.7-Max, a new model designed for long-horizon autonomous agentic tasks. The model demonstrated up to 35 hours of continuous autonomous execution during an engineering task and features a 1 million token context window. Qwen3.7-Max is a proprietary model that outperforms all Chinese competitors on reasoning and coding benchmarks and matches Claude Opus 4.6.

OpenAI updated Codex with richer context, goal mode, browser improvements, and locked computer use. The latest Codex release added Appshots for attaching macOS app windows to Codex threads, general availability of Goal Mode across the Codex app, IDE extension and CLI, improved browser annotations, and locked computer use for eligible Mac Computer Use users. The update is aimed at making Codex more useful for longer-running software.

Cohere unveiled Command A+, a highly optimized 218B parameter language model released under a permissive Apache 2.0 license for open-source enterprise use. The Command A+ model utilizes a Sparse Mixture-of-Experts architecture and key features include hardware-efficient quantization for single-GPU deployment, multimodal capabilities, and improved tokenization for non-European languages. It is available on Hugging Face.

Amazon Nova Act now qualifies as a HIPAA eligible service. This expansion allows healthcare organizations to deploy autonomous, browser-based AI agents to automate complex workflows involving protected health information. The service can automate tasks such as appointment scheduling, insurance verification, and claims processing.

Anthropic announced updates to Project Glasswing, allowing qualifying customers access to a Claude harness, a threat model builder, and various skills. The company also plans to expand the project to additional partners and has released a dashboard for open-source vulnerabilities.

Cerebras Systems announced high-speed inference for the trillion-parameter Kimi K2.6 model. The chipmaker is running Moonshot AI’s open-weight model at 981 output tokens per second, significantly outperforming GPU-based cloud providers. This enterprise-first deployment utilizes wafer-scale architecture to provide massive speed improvements for agentic coding and heavy workloads.

Copenhagen-based healthcare AI Corti is launching Symphony for Speech-to-Text, a new generation of clinical-grade speech recognition models. The models achieved a 1.4% word error rate on English medical terminology, significantly outperforming generalist APIs from OpenAI, ElevenLabs, and Whisper. The technology also demonstrated a 98.3% recall rate on clinical entities and surpassed the performance of the legacy incumbent, Dragon Medical One.

AI Research News

An OpenAI reasoning model autonomously disproved a major conjecture in discrete geometry called the unit distance problem. OpenAI said an internal general-purpose reasoning model produced a proof resolving the long-running planar unit distance problem, originally posed by Paul Erdős in 1946. The proof, reviewed by external mathematicians, is notable because the model was not a math-specialized system and used ideas from algebraic number theory to disprove a conjecture many mathematicians believed was likely true.

This marks the first time AI has autonomously solved a prominent open problem central to a field of mathematics. … The result is also notable for how it was found. The proof came from a new general-purpose reasoning model, rather than from a system trained specifically for mathematics, scaffolded to search through proof strategies, or targeted at the unit distance problem in particular.

ComplexMCP was introduced as a benchmark for LLM agents in realistic tool-use environments. The paper on ComplexMCP argues that many agents can call isolated APIs but struggle when tools are interdependent, noisy, and embedded in workflows that resemble commercial software automation. The benchmark is intended to measure the “last mile” of agent performance, where success depends not just on calling tools but on managing state, dependencies, and changing environments.

A new benchmark called SMDD-Bench tests whether LLMs can solve real-world small-molecule drug discovery tasks. The benchmark evaluates frontier open and closed models on tasks requiring chemical and biological reasoning, 3D intuition, specialized tool use, and planning under limited oracle calls.

The paper “Benchmarking Autonomous Agents against Temporal, Spatial, and Semantic Evasions” found that autonomous agents such as OpenClaw are vulnerable to systemic, architecture-level vulnerabilities that exploit multi-turn interactions. The study evaluated a standard agent framework with 10 mainstream LLM backbones across 20 threat scenarios and found that an evasion framework raised the average risk trigger rate from 28.3% to 52.6%. They constructed A3S-Bench to evaluate AI agents on such vulnerabilities.

AI Business and Policy

Andrej Karpathy has joined Anthropic as an individual contributor on the pre-training team. This is big news because he is a highly-respected AI researcher, one of the pioneers in the AI space, who cofounded OpenAI and led AI at Tesla for many years. He’s the 3rd senior former OpenAI employee to join Anthropic in the last two years.

Nvidia reported record first-quarter fiscal 2027 revenue amid increasing AI infrastructure demand. Revenue for the quarter ending in April reached $81.6 billion, up 20% from the previous quarter and 85% from a year earlier. AI compute demand from hyperscalers, enterprises, and AI labs to support training and inference continues to expand.

Google and Blackstone announced a new AI cloud infrastructure venture that will serve up Google TPU AI support through a compute-as-a-service model. Blackstone will initially invest $5 billion in equity to help bring 500 megawatts of data center capacity online by 2027, and total investment could reach $25 billion including leverage.

The White House prepared an AI oversight executive order focused on model review and cybersecurity risks that would create a voluntary framework for AI labs to engage with the Federal Government before releasing covered models. The proposed order would task agencies with evaluating AI security following concerns regarding models like Anthropic’s Mythos and OpenAI’s GPT-5.5 Cyber.

However, Trump called off those plans because of concerns such regulation could dull America’s edge on AI technology. A major point of contention is a requirement for companies to share advanced models with the government up to 90 days ahead of launch. The turnabout reflects tension between AI safety advocates pressing for stronger regulation and tech-industry allies who favor voluntary cooperation.

U.S. lawmakers have moved to counter Chinese AI and technology exports with legislation to bolster American exports, as part of a broader geopolitical contest over AI infrastructure, chips, and digital technology.

OpenAI and Dell Technologies are collaborating to deploy Codex in the enterprise, integrating Codex with the Dell AI Data Platform to support hybrid and on-premises workloads. Codex-powered agents will be utilized for tasks including software development and business workflow automations.

Regis Aged Care implemented RegiCare Assist AI assistant to streamline clinical care management. Developed with Microsoft Copilot Studio and Microsoft Foundry, the assistant summarizes voluminous progress notes and flags clinical concerns.

Spotify and Universal Music Group have entered a licensing deal to launch an AI remix tool that allows fans to create AI-powered covers from UMG’s catalog. Spotify also revealed new AI-driven tools for audiobook and podcast production during its Investor Day.

AI Opinions and Articles

Axios reported this week on some fresh polling on AI policy, and the results show Democrats becoming more AI skeptical, Republicans more likely to trust AI companies, and anxiety among the younger voters that AI will harm job opportunity. Overall, a majority are “pro AI innovation” and have a nuanced view of AI regulation, between the poles of “Pause AI” and “get Govt out of the way.” The most popular position - a 41% plurality - is that the government needs “basic safety” standards that keep American companies competitive:

In a twist of irony, Steven Rosenbaum explains how inaccurate quotes got into his book The Future of Truth. A New York Times investigation revealed that the use of AI tools during research led to several improperly attributed or synthetic quotes in the book. Rosenbaum is currently conducting a citation audit to correct these errors in future editions. The book is about “how Truth is being bent, blurred, and synthesized” thanks to the “pressure of fast-moving, profit-driven AI.”

He blames AI because it was deceptively easy to use, but he proves that AI slop is due to sloppy fact-checking and editing and humans cutting corners. A good workman never blames his tools.