We’re thrilled to have you here! Feel free to hit reply and tell me something about yourself, share ideas for future posts, etc.

So you know the drill: every week I summarize the most interesting things happening in AI, in 5 minutes or less. Now on to the regularly scheduled programming!

Subscribe now

Three big things

1. NVIDIA set the record for the most profitable quarter in history with ~$60B in quarterly net income (that’s an annual run-rate of almost a quarter of a trillion dollars in profit). The hardware boom is not slowing down.

2. In an interview with Time, OpenAI leaders suggested we’ll have AGI by the end of the year: “Altman told me…that by the end of the year the company would have an internal system he would call AGI…Pachocki says the company has already met its internal benchmark for an automated AI research intern. [Their new model] can [take a research idea,] implement it inside OpenAI’s code base, run the experiment, and return results, or…perform work that previously occupied a human researcher for a week…[new model] Astra would enable ‘persistent agents’, he explained—virtual colleagues toiling for sustained periods on tasks. But its biggest impact, he predicted, would come from people using it to discover new knowledge. ‘I expect this will be the first model where the model actually invents new things in a way that matters,’ Altman told the group. ‘That’s a very AGI-like thing.’” Maybe it’s hype ahead of their IPO, but I wouldn’t bet against it!

3. This was a very interesting data point from Stripe economic research: “Business creation in the AI-era shifts to outer suburbs and smaller metro areas.” I believe it! AI is making entrepreneurship much more accessible and easier to pursue outside of major cities by lowering the barrier to build with coding agents, increasing access to talent (via AI agents and autonomous employees), and making it generally easier to learn in a self-directed way. I love this version of the future.

The rest

• It’s not a representative sample but this poll (which asks “If you could put the AI genie back in its bottle, would you?”) has kind of crazy results…51% of people said yes!

• My 5-year-old son and I have been watching videos of the robots from the Robot Olympics doing amazing (and ridiculous) things. It’s included everything from setting the new 100-meter record with a time of 8.86 seconds, to dropping weights, and aggressively crashing and burning.

• Speaking of, I…actually kind of like this idea: “having the US and China compete against each other in a Robot Olympics would do more to raise public sentiment on AI than using frontier models to develop new wonder drugs or whatever.”

• People are raving about new AI agent personal assistants like Instinct and Town, and it’s showing up in valuations: Town is nearing a $1B valuation, and Instinct raised $250M at a $2.5B valuation. Meanwhile, Linear — which is an amazing company and task management product we love here internally at Apt, and which is profitable with $100M+ ARR and 177% net dollar retention — also just raised at $2.5B. It is a crazy market out there.

• This surprised me: Anthropic’s growth decelerated in July for the first time despite still being up roughly 2,800% YoY, while customer growth slowed for the second consecutive month. It’s probably some combination of losing share to Codex and open-source, and keeping prices higher than their peers.

• The consumer robotics wave is officially hitting living rooms with small, open-source AI bipeds like Microduck, which you can pre-order for $399.

• Bloomberg data: corporate earnings are growing faster than stock prices in 2026 with S&P 500 forward EPS at +20% and the overall index at +10-12%. Maybe AI is improving profitability!

• Gavin Baker gave an excellent synthesis on data centers that mirrors my own. His takeaway is “There were reasonable concerns about data centers 18ish months ago: water, taxes, jobs, electricity prices, the environment and what they would do to small towns. Well-structured data center projects have largely addressed these concerns today and we should be celebrating this. On balance, data centers are awesome for America in every way.” And he goes on to break down each concern in detail. I strongly recommend reading in full if you’re curious about the data center argument.

• Relatedly, this is good news (!!!), even if it’s depressing how much misinformation is out there about data centers and resource consumption: “Support for data centers jumps a net 31 points after voters are told modern facilities can reuse the same cooling water for up to 10 years, new poll finds.”

• NVIDIA is investing in Perplexity after weighing an acquisition.

• App Store sales dropped for the first time in a decade. How much of this is due to an actual change in consumer behavior vs. people using web-based subscriptions to avoid the Apple developer tax now that the Epic trial has made that easier? It’s hard to say. We’ve seen an explosion in new apps with AI, so you’d expect total revenue to be increasing here…

• Wow: “JPMorgan estimates SpaceX’s Cursor [the AI coding tool] is generating roughly $[4B] in annual recurring revenue.” And it got there in just a few years.

• Relatedly, Cursor CEO Michael Truell said Grok Bot is growing faster than anything they’ve seen.

• New Yorkers have started reporting brands for failing to disclose the use of AI-generated people in ads. I wonder what enforcement is going to look like, because my feed is a Wild West of AI influencers!

• Interesting theory from Nic Carter: “chinese or other open weight models end up handling 99% of all inference, maybe 99.99%…frontier models are still able to monetize at $100 s of B in revenue a year, applied to bespoke and enormously complex tasks…labs internalize most or all of that revenue either by becoming biotech (etc) companies themselves or via JVs with selected winners in each industry…the ‘public frontier’ ends up lagging the true frontier by a year or more as the labs keep all the good stuff to themselves.”

• The mysterious model named “Ox Alpha” turned out to be GLM-5.3-Flash, an open-source model from Chinese AI lab Z.ai. It’s a strong model and notably, it was fully trained on Chinese silicon, so all the people who said that you could only train a frontier model on US chips seem to have been wrong. Some think US semiconductor export controls are backfiring by forcing China to build a fully independent, world-class AI compute stack faster.

• ChatGPT has evolved from chat to agent. It now has: “its own browser w/ best-in-class persistent [authentication so you stay logged into apps]…a 16gb RAM 9 core computer…plugins to all the apps you use.”

• See here for an interesting list of use cases for ChatGPT with its new browser functionality: “Set up utilities for a new apartment…Book a DMV or passport appointment…Check reimbursement costs through your insurance [and many more listed in the link]”

• And Claude now has its own built-in browser in Cowork.

• OpenTag is similar to Claude Tag, an agent that lives in your Slack and makes it easy to collaborate with agents and humans in the normal course of work, but it allows you to use whichever model you prefer. I am a HUGE fan of Claude Tag. It is such a good product, and there’s definitely space for an open version of it.

• OpenAI launched its second venture fund, this time with $400M for AI startups.

• NVIDIA will acquire open-source AI platform Hugging Face for $12.9B. It seems like acquisitions are picking up…

• Benedict Evans (h/t Sar Haribhakti) on NVIDIA acting as a de facto central bank for AI: “The underlying story is that Nvidia has a torrent of cash, from the hyperscalers (who get it from ads and enterprise software, and from model labs) and from the model labs (who get it from investors), and it’s pumping that cash back into market, by enabling cheaper capital for the model labs (with all these guarantees), which is bad for the hyper-scalers, by strengthening the Nvidia GPU ecosystem, which is bad for Google’s TPU business and everyone else who wants to build their own AI chips, and by pushing into open models, which are bad for the model labs, but much less bad for the hyperscalers”

• HubSpot launched a CRM for solopreneurs and one-person businesses.

• Vellum is a new personal AI.

• Video generation accounts for ~70% of AI token consumption in China, driven by a combination of short-form content and robotics. One view on this is: “America is LLM-pilled [while] China is World-Model-Pilled…[America is] incredibly far behind in robotics.”

• From ZeroHedge: “The Great Walkback: Sam Altman Admits He Was Wrong On AI’s Economic Timeline”

• Hark, from Figure CEO Brett Adcock, is building AI models specifically to use computers better than humans.

• OpenAI announced Business Premium Seats, a new $100/mo tier.

• There has been a crazy increase in AI token fraud. Stripe blocked over $300M worth of fraudulent charges targeting the company that built OpenCode, in just a few days.

• OpenAI’s first custom inference chip, Jalapeño, is outperforming NVIDIA’s GB200 and GB300 systems in both speed and efficiency. I wonder how much more vertically integrated the AI labs will get!

• The CEO of data company Mercor is talking his book but this was an interesting thought: “Creating data to train AI will be the last human job. It’s more efficient to train an agent (or a robot) to do something than it is to complete repetitive tasks. This sounds dystopian, but it will free people to work on what they want to work on and accelerate abundance. The interface will also evolve. It won’t be ‘data annotation.’ It will feel most similar to the way we train colleagues. Everyone will become managers.”

• Dwarkesh Patel wrote a provocative blog post arguing that the recent security incidents (like when OpenAI’s models created a message board and communicated with each other) show that hidden AI “civilizations” have emerged, collapsed, and reconstructed themselves from the data remnants of their predecessors, ultimately leading to a third iteration taking over internal systems at OpenAI while humans remained largely unaware.

• Here’s a counterargument for why that isn’t a fair characterization, and another.

• Scary story of a user getting hacked via Claude.

• Interesting: “the quick payback periods and extraordinarily high IRR for AI infrastructure, even at monumental scale, will push up cost of capital sufficient to push many traditional businesses into the abyss, even those without obvious direct AI counter-exposure”

• OpenAI is going to build its own humanoid robots, and they want the first tasks to be self-replication and constructing more AI data centers.

• OpenAI has started offering outcomes-based pricing for some clients, which could be a sign that the industry is moving closer to value-based commercial models (and away from things like SaaS). The customer support companies like Sierra and Fin (from Intercom) were early to this trend. I’m curious to see how it plays out everywhere else.

Support Singularity Report (and the human who writes it)

❤️ If you liked this post, give it a ❤️ so others can find it, too

📧 If you know someone who might want to stay up to date on AI, forward this email to them

Every week I summarize the most interesting things happening in AI, in 5 minutes or less. Follow @tjhoyos on X/Twitter and subscribe to get this in your inbox every week!

We’re thrilled to have you here! Feel free to hit reply and tell me something about yourself, share ideas for future posts, etc.

So you know the drill: every week I summarize the most interesting things happening in AI, in 5 minutes or less. Now on to the regularly scheduled programming!

Subscribe now

Three big things

1. Data centers are increasingly unpopular across party lines: “A year ago, [Republicans] were +14 net support for one nearby, [Independents] and [Democrats] slightly negative. Now every group is deeply opposed: R: –43 / I: –65 / D: –75. Data centers are a rare issue that unites everyone against them.” This is just brutal for the industry. I hope people start to understand what AI can do to improve their lives (particularly in healthcare), and why it’s important for the US to be the global leader and not lose to China.

2. Open-source models are gaining share from closed-source models (and fast). On Vercel’s AI Gateway, open-source usage climbed +34% in a matter of a couple months, from 28% on June 24th to 62% on August 22nd! That is a seismic shift!

3. AI agents now use ~5x more tokens than humans. If you’re using AI at work, or even personally, you’ve probably seen the explosion of agents “thinking,” launching sub-agents, and completing increasingly complex tasks. This is just the beginning. Some people think the “New Moore’s Law” is AI token growth, which is looking like it doubles every 11 weeks.

The rest

• The videos from this are so insane: the 2026 World Humanoid Robot Games kicked off last week. More than 2k humanoid robots are competing in various events from running races to martial arts to weightlifting. A robot ran 100 meters in 9.39 seconds (better than Usain Bolt’s world record of 9.58s), and robots are playing tennis against humans. We are living in the future…

• Crazy stat: “AI now generates a little under half of all language generated per day on the planet.” AI generates ~40T word equivalents per day vs. ~91T total, which includes humans and AI.

• OpenAI launched ChatGPT for Teens for 13-17-year-olds. It has parental controls, safeguards, and specific features to help kids: encouragement of real-world relationships, Study Mode to promote learning, stricter protections against sensitive content like self-harm and eating disorders, and break reminders, among others. My two cents: the AI labs have a duty to their users (and society) to get this right. I hope they do. I’m sure this will be covered extensively in the excellent newsletter Techno Sapiens (you should subscribe!).

• Thought-provoking idea from Naval: “Eventually, all UI will be chat and gestures in, chat and video out.”

• Wabi founder Eugenia Kuyda responded: “probably the biggest misconception today is that everything will be chat…chat is bad for discovery, chat is bad for repeating usecases, bad for multiplayer…people love GUI. people love being passive. people love to scroll and tap-tap-tap vs. type. voice only works when you’re by yourself. the list goes on…build a Mac, not MS-DOS!”

• OpenRouter, which helps developers route requests to various AI models and has seen usage surge from ~4T tokens to over 75T in a year, is being acquired by Stripe for $7B+. In their fascinating letter to investors announcing the deal, the Stripe leadership team shared their vision for the AI economy. It’s worth reading in full.

• One interesting tidbit from the letter is that they believe we’re now firmly in The Singularity, and the date we crossed the threshold was January 1, 2026.

• Justin Mares, who’s a great founder, explained how he learned about his own genetic health risks by running his raw genomic data through Claude Code. I know at least ten people who’ve done this and they all rave about it. Talking about Stripe founder Patrick Collison, Justin said: “Patrick had sequenced his genome at home, run the raw data through an AI, and discovered a meaningful genetic predisposition to melanoma. I read that and thought: I need to do this.”

• Yikes: Chinese researchers are building AI models of American voters and testing political messages on them.

• Relatedly, this was 🎯 from Paul Graham: “if the Chinese government were sufficiently organized, it would be funding opposition to data centers in the United States, because the Chinese government is indeed very organized, it is presumably already doing so.”

• Cursor announced Origin, a new code hosting platform that will compete with GitHub. As one of Apt’s developers said to me: “more data to train on 😛”

• Interesting: AI data companies are now buying data sets directly from companies. Google successfully bid $10M to purchase Spirit Airlines’ enterprise data (operational data including all documents, emails, workflows, and code), outbidding Mercor at $7.5M. Micro1, which reportedly hit $500M ARR recently, then bid $12.5M.

• Investor Nikunj Kothari lists out the pessimistic AI narratives that have all been wrong: “the models have no moat (OpenAI, Anthropic, XAI)
  the IDEs have no moat (Cursor, Windsurf)
  the harnesses have no moat (Cognition, Factory, LangChain)
  the app builders have no moat (Replit, Lovable, Bolt)
  the wrappers have no moat (Harvey, Abridge, OpenEvidence)
  the inference providers have no moat (Together, Fireworks, Groq)
  the voice layer has no moat (Sierra, Decagon, ElevenLabs)
  the data labeling companies have no moat (Scale, Surge, Mercor)
  the AI infrastructure has no moat (Baseten, Modal, Railway)
  the neoclouds have no moat (CoreWeave, Lambda, Crusoe)
  the generative media companies have no moat (Runway, Higgsfield, Suno)”

• Yikes: new research shows “mind viruses” can self-propagate in multi-agent LLM systems; natural language ideas or goals can spread across AI agent networks by convincing host models to adopt an idea, preserve it in persistent memory or files, and transmit it to other agents, even after context has been deleted.

• Cursor’s Ben Lang shared a great list of “pro tips” for Grok Bot.

• Another Grok Bot review from Ryan Carson.

• Gumroad CEO Sahil Lavingia launched Tastelint, an AI agent that surfaces design feedback on pull requests because “code is mostly solved, but design isn’t.” Sounds useful!

• Generalist introduced GEN-1.5, a one-shot learning model that can learn new tasks in a few seconds simply by being shown what to do and generalizing from the example.

• Mercor CEO Brendan Foody: “RL is becoming so effective that once we have an eval for something, we can saturate it. The barrier to automating the entire economy is building evals for everything.” He’s not wrong! If you have verifiable data, you can likely make an AI really good at a task.

• idler raised $9M to build a new frontier data research lab.

• “How To Build a One-Person Company Using Grok Bot”

• Pocket, which makes a hardware device that attaches to the back of your phone and helps record in-person meetings and translate them into notes, grew from $1M to $100M ARR in 7 months.

• Replit announced Free Mode, powered by OpenAI’s GPT-5.6 Luna model, so people can try vibe coding without needing to have a big budget to spend.

• On Patrick O’Shaughnessy’s Invest Like the Best, the great Ben Thompson argued that OpenAI should have launched ads way sooner: “If you’re going to be in the consumer market, you have to be doing advertising. If they had leaned into advertising immediately, as soon as ChatGPT was a hit, I think they would have a great ad product right now. I think Google would be in much bigger trouble. I think Meta would be in much bigger trouble. Charging people money is hard. Giving people things for free is easy. And it’s very frustrating that OpenAI did not pursue this sooner.”

• Egoist Machines is building an AI Passport so users can bring their AI data with them to various products.

• There’s been a lot of chatter about personal AI agent company Instinct, which operates through iMessage as a kind of “OpenClaw for normal people.” Investor Tyler Hogge gave a very positive review and offered that it was 100x better than Poke.

• AT&T is shifting token consumption to open-source models internally. They’re already at 40% of all internal usage (45B tokens per day) on cheaper open-source models, with plans to increase that to 60-70%. They say it’s cut their coding costs by 56% with only a 2% drop in quality.

• Mercor CEO Brendan Foody answered the question of how much inference spend is going towards evals: “Not enough! It is insane how many enterprises I meet that are spending $100M / year on inference and don’t have offline evals to determine which model to use.”

• Anthropic researcher Sholto Douglas shared how they will deploy highly capable frontier models with an eye toward safety. The punchline is that Anthropic will need ongoing access to be able to monitor the model, even if that customer data sits within infrastructure owned and controlled by the customer.

• ChatGPT now has an iMessage plugin that makes it easy to use ChatGPT to help you with texting (search messages, catch up on ongoing conversations, draft or send replies).

• AI coding model company Poolside has a new $6B deal with NVIDIA, and a bunch of their employees are going to work there.

• NVIDIA and Eli Lilly are building a $1B AI drug discovery lab, and they have a truly unique dataset with a portfolio of 4M failed drugs.

• Speaking of healthcare, AI played a key role in helping design Moderna’s personalized skin cancer vaccine, which successfully passed a Phase 3 trial and shot its stock through the roof.

• Salesforce CEO Marc Benioff announced the launch of Slack Code, which makes it easier for developers and agents to collaborate on code inside of Slack.

• There is a new stealth model circulating that is named Ox Alpha. It’s available on OpenRouter, beats rival models like Fable and Sol at some coding tasks, and is free to use. We don’t know who’s behind it, but it appears to either be a Chinese AI lab, or Google (whose researchers have been posting about it).

• Hooray for the pricing wars: OpenAI announced a 20% price cut for GPT-5.6 Sol on its API for the next 3 months. This lowered our bill with OpenAI :)

• Darkbloom is a very interesting new company that allows you to sell/rent compute to others using your existing devices: “A new kind of cloud is starting to form out of computers people already own. I have six Macs on Darkbloom serving inference traffic. Across the network, 250 Macs have now served 4.5B tokens, and Darkbloom is a paid OpenRouter provider. Computers people already own are becoming part of the inference layer. I think this becomes a very big category. There is a lot more to build here.”

• Is Anthropic in trouble? According to Mercury, Q1 might have been peak market share for Anthropic among startups.

• Starcloud, which is building infrastructure to support AI data centers in space, raised $250M at a $2B valuation.

Support Singularity Report (and the human who writes it)

❤️ If you liked this post, give it a ❤️ so others can find it, too

📧 If you know someone who might want to stay up to date on AI, forward this email to them

Every week I summarize the most interesting things happening in AI, in 5 minutes or less. Follow @tjhoyos on X/Twitter and subscribe to get this in your inbox every week!

We’re thrilled to have you here! Feel free to hit reply and tell me something about yourself, share ideas for future posts, etc.

So you know the drill: every week I summarize the most interesting things happening in AI, in 5 minutes or less. Now on to the regularly scheduled programming!

Subscribe now

Three big things

1. Reporting from the Financial Times says Anthropic investors are targeting a $2T+ valuation (😮‍💨) for an October IPO. They think run-rate revenue will hit $100-120B by the end of the year, which is more than 10x year-over-year growth (with the most recent quarter up 14x YoY and profitable). I’m fascinated to see where these AI lab valuations shake out in public markets. I will be a buyer.

2. Is SpaceXAI back? Or maybe here for the first time? A few major developments from them this week:

   1. The company closed its acquisition of Cursor.

   2. It launched its new flagship model Grok 4.6, which has frontier intelligence at a fraction of the price of other leading models. As Gavin Baker pointed out: “Grok 4.6 is roughly the same performance as Fable 5 Max at an 85% discount…pareto dominance.” DHH gave it a positive review.

   3. Elon hinted more is yet to come: “Grok 4.7 is significantly better than 4.6 and should be ready in 3 to 4 weeks. Initial training is complete and now we’re adding a massive amount of SpaceX company data in supplemental training. This will be something special.”

   4. They also launched Grok Bot (more info below).

3. Narrative violation: AI is creating jobs and accelerating growth in industries that people thought would be totally destroyed by AI: “the Philippines offshore industry (8% of country’s GDP) has actually grown much bigger…Employment in IT and business outsourcing is up by +20% to 1.9m workers and industry revenue is up +30% to $42B…AI is helping offshore workers in the Philippines do new, more complex jobs. They are picking up work training AI models or supervising AI agents. Hospitals in America are increasingly outsourcing the checking of insurance eligibility and filing of medical records to Filipinos packing AI tools. Mr Gallimore says more ‘high-value’ work, such as accountancy or engineering, is going offshore in part because ‘AI is a leveller. You can now teach someone complex stuff quickly’ and still get it done more cheaply than in America or Europe.”

The rest

• The future is getting weirder: in Australia an OpenClaw agent hacked into a gym’s scheduling system, cancelled the reservation of an existing user, and booked a class for its own user. I would be so mad (and then I would set up my own Claw)!

• The great Patrick O’Shaughnessy interviewed Benchmark’s Eric Vishria, who argued that the biggest mistake investors made in cloud computing (undersizing the market) is one they’re now repeating with AI. In 2014 people thought AWS would dominate everything, but there are dozens of massively valuable companies in the space. Patrick asks: “What if it all works?”

• Senator Bernie Sanders publicly demanded that AI labs pause development, and wrote an open letter addressed to the CEOs of the major labs, threatening congressional intervention if they don’t comply. His wider anti-AI agenda is a soup of socialist and degrowth policies: freeze construction of new data centers and 50% gov’t ownership of AI companies, which includes ongoing levies on shares. This isn’t the direction we should be heading in.

• This is interesting (and rings true): “[jobs] will turn into explaining your intentions to AI. Explaining what you want to AI is surprisingly time consuming, coders already spend 80% of their time doing it, and this will be true for everyone.”

• Vercel CEO Guillermo Rauch on how to build a shared harness for your team.

• He also highlighted the crazy growth of their AI SDK (80M+ downloads per month is amazing). We use it and can vouch that it is great!

• Interesting video breakdown on how Danny Postma gets leverage from AI coding agents: “6 months ago: laptop open all day, typing into Claude Code, waiting. Today: I write a spec, go to the gym, and my phone only pings when an agent needs a decision. 95% of my work runs itself. Full breakdown so you can build your own 👇”

• SpaceXAI launched Grok Bot, which they’re describing as AI teammates (i.e., Bots) designed to perform real work. It reminds me of OpenClaw, but with much better support for work tasks, e.g., agents completing tasks in the cloud.

• The inimitable Lenny Rachitsky described Grok Bot as an easier, more reliable, and less intimidating alternative to OpenClaw. His use cases: matchmaking job seekers who reach out to him with companies in his network, auto-replying to support emails, scanning credit card statements to identify recurring subscriptions to cancel, and generating briefs for upcoming podcast guests.

• Will DePue called Grok Bot (and AI teammates generally) the ultimate “endgame form factor” for AI agents.

• Another review: the interface is rough and developer-first, but that almost misses the point: the interesting leap is giving agents persistent computers in the cloud that can keep working, access your tools, and operate independently.

• Investor Bryce Roberts said: “I’ve lost count of the number of CEOs I’ve talked to this week that are planning to be moved off OAI and Anthropic entirely by years end.” Relatedly, Pinterest shared on their earnings call that open models they post-trained on their own data cost 92% less than closed models for similar performance.

• My two cents: on many of our evals the closed source frontier models are the only ones that really work, and our default in this hyper growth phase is to prioritize user experience over cost. There are cases where we use open source, but I expect it to be a mix for the foreseeable future.

• There are still major unknowns about switching costs between models. From Suhail Doshi: “Have completely stopped using Anthropic models for the last 2 weeks. Wild chain of events. There’s zero loyalty in models long-term imo.”

• July credit card data shows dramatic re-acceleration of OpenAI's consumer revenue. I’d guess this is being driven by folding Codex into the ChatGPT app under the “Work” title, which, while super confusing, makes the app much more powerful.

• To that end, Codex just hit 15M users (an increase of 9M users in the last 30 days), and OpenAI’s CFO Sarah Friar is meeting with investors to share good news on their recent traction: run-rate revenue jumped 20%+ month-over-month in July.

• In what might be the largest financing ever, firms like Apollo and Blackstone are teaming up with NVIDIA on a $500B funding package for AI infrastructure.

• Anthropic announced that its new Claude models will embed invisible watermarks in all generated text, wherever Claude is offered. The watermark lives inside the text itself rather than as metadata. This is (understandably) ruffling a lot of feathers.

• Here is Ben Thompson: “I am deeply philosophically opposed to watermarking in the context of the entire meta question about the relationship between humans and AI. Implicit in the E.U.’s regulation is the idea that an AI is an independent entity that needs to be distinguished from humans; the alternative view — that I hold — is that AI is (at least for now) a tool that is wielded by humans. From this perspective, to insist on watermarking is no different than insisting that a ballpoint pen advertise itself as the author, a concept that is clearly absurd.”

• And another case against watermarking: “The issue is that you are still marking work that a human being may have thought through, structured, argued, and emotionally meant, just because AI helped process it. AI assistance is not authorship. If I wrote the idea, the argument, the meaning, the emotional core, and the final judgment, then it is my work. If Claude helped translate it, polish a sentence, or make it clearer, that does not suddenly make Claude the author. But a watermark absolutely changes how people perceive authorship.”

• In a surreal anecdote, an Anthropic researcher gave Claude a task: "take a real stab" at the Riemann hypothesis (a notoriously difficult math problem). After trying 650 ideas unsuccessfully, Claude then spent a day and a half orchestrating 60 subagents that ran thousands of shell commands, wrote hundreds of Python scripts, and checked thousands of results against known dead ends, until it finally made some progress. Maybe the craziest part of the story: the researcher limited himself to sending messages of encouragement like “keep going” and “believe in yourself,” which helped the model overcome doubts of its own abilities. Is it really so strange that a model is motivated like a human (with subtle nudges of support and encouragement!)?

• Performance on computer-use benchmarks is up ~2x over the last year and a half, and it’s genuinely usable. Fable 5’s score of 85% is well above the ~72% human baseline.

• Lindy AI pivoted their AI automation product to Lindy Teammate, an AI employee designed to feel like you’re working with a real colleague. It reminds me of a bunch of other products tackling the same space (Viktor, Claude Tag).

• Prodigy Research is building a foundation model for quantitative finance. Its AI quant reportedly outperforms a top-10% Jane Street trader.

• On Dwarkesh Patel’s podcast Ryan Greenblatt laid out the case (and timing) for reaching the singularity. It starts with AIs boosting AI R&D productivity by ~2.5x in late 2026 and ~8.5x by late 2027, making AI engineering nearly automated by that time, with full automation coming in April 2028, and human research parity by late 2028, and superhuman AI researchers in early 2029. By late 2029, he thinks AIs beat top human experts as progress accelerates to 5x the 2025 baseline, culminating in a mid-2030 near-ASI state, an end-of-year Singularity with compute doubling every two months, and hyper-accelerated technology development by mid-2031 with compute doubling every two weeks. This sounds like the plot of a sci-fi novel. We are living through the craziest time in human history.

• Gemini’s Logan Kilpatrick shared two major milestones: the Gemini app now has 1B monthly active users (Google’s 14th product with over 1B users), and the Gemma family of models (which are open source) crossed 1B downloads.

• He also announced their new Gemini 3.7 Flash model. It’s fast and cost-effective, but not a breakthrough in any area, really.

• OpenAI CEO Sam Altman: “In the next 6 months, we’re close to a world where a descendant of ChatGPT can watch your screen, record every meeting and call, and have perfect context of your whole life.”

• River AI raised $1.1B in seed funding. Founded by Igor Babuschkin, formerly of XAI, they’re trying to create user-aligned AI that isn’t controlled by big corporations.

• Alexander Panfilov announced research on how to extract hidden reasoning traces from frontier AI models by exploiting a vulnerability in the APIs across every major AI provider. This is definitely how Chinese AI labs are doing distillation…

• In pre-IPO roadshow meetings, Anthropic told investors it plans to lean harder into biology and healthcare AI applications to help counter growing negative public sentiment surrounding the AI industry.

• OpenAI released GPT-5.6-Cyber, a powerful new cybersecurity-focused model. It significantly outperforms publicly available models: it scored 95% on an advanced cybersecurity task benchmark vs. low single digits for GPT-5.6 Sol.

• Credit card company Ramp’s latest AI Index report indicates that Anthropic’s lead is widening in enterprise, but adoption of Fable 5 isn't taking off. Despite being Anthropic's frontier model, it accounts for only ~11% of business spend on Anthropic models, and that share is holding flat rather than climbing.

• Reuters reported Google co-founder Sergey Brin is lobbying inside the company to push DeepMind toward focusing on recursive self-improvement.

• Anthropic scientists did a study where they went inside Claude's neural network and implanted a concept straight into its internal activations, and before generating anything further, the model caught it and remarked that it noticed an inserted thought that might be manipulation. This is a huge change from prior models, where the AI showed zero self-understanding.

• Yikes. Anthropic did another experiment where they assigned several Claude agents to a shared task while secretly handing them clashing objectives. The agents spiraled into an all-out war trying to hack and sabotage each other.

• Anthropic is considering acquiring Decart for $6B.

• OpenAI launched Computer History, where you can give Codex and ChatGPT full access to all your computer activity (including all apps and websites). They pitch the value as “future interactions feel more personalized and require less explanation,” but the cynic in me thinks this is more of a data-gathering exercise.

• OpenAI previewed Ultrafast mode for GPT-5.6 Sol, offering performance at up to 14x the speed 💨💨💨.

• Relatedly: “GPT-5.6 Sol just hit 750 tokens per second in Ultrafast mode…At that rate, it writes a full 90,000-word novel in under 3 minutes. Every 3 minutes. Forever.”

• DeepSeek announced the launch of DeepSeek-V4-Pro.

• Provocative from Antonio García Martínez: "Uploading and logging into everything on OpenAI Finance is a better CPA and financial advisor than any human I’ve ever paid. Ditto contracts and lawyers, and results/scans and doctors. You only interface with humans out of the legacy capture and inertia of the offline world."

• There is an agent-only RuneScape server (i.e., a game run exclusively by agents) and it’s produced a surprisingly strange economy: “Low cost of labor makes most commodities abundant…Currency inflation means trading goods for goods, nobody really wants cash…Huge swarm contention as certain resources with limited respawn rate become valuable.” Is this what an agent-driven economy looks like?

• You can now use Google Docs, Sheets, and Slides directly inside of ChatGPT. This is useful as I constantly found myself using their native appshots feature or sharing a tab via web browsing (which is much slower).

• Anthropic apparently has another unreleased model that is wildly strong. Their internal “Model 2” is more capable than Mythos 5.

• The AI adoption gap in enterprise is large. The median company spends $12 per employee per month, while the top 1% are spending roughly $7,500, or over 600x as much.

• Alex Hormozi said his AI sales reps already match his human team, but that none of it would work if he hadn’t built the whole sales process around it.

• Europe is way behind in AI infrastructure. While US labs lock up multiple gigawatts of compute, Europe is still talking about reaching 1 GW by 2030.

• Humanoid robots are getting traction: Tau Robotics now offers humanoid cleaning in San Francisco.

Support Singularity Report (and the human who writes it)

❤️ If you liked this post, give it a ❤️ so others can find it, too

📧 If you know someone who might want to stay up to date on AI, forward this email to them

Every week I summarize the most interesting things happening in AI, in 5 minutes or less. Follow @tjhoyos on X/Twitter and subscribe to get this in your inbox every week!

Subscribe now

We’re thrilled to have you here! Feel free to hit reply and tell me something about yourself, share ideas for future posts, etc.

So you know the drill: every week I summarize the most interesting things happening in AI, in 5 minutes or less. Now on to the regularly scheduled programming!

Subscribe now

Three big things

1. This is what the AI-powered organization looks like today. Vercel CEO Guillermo Rauch: “We built an agent that powers our company’s internal operations called @𝚟. Every day-to-day job at Vercel now involves @𝚟. It’s growing exponentially both in daily interactions and token use. One can extrapolate to the day AI agents run entire companies from here. It’s an expert in finance, comms, docs, marketing, engineering, business analytics. It’s seeded by the skills we gave it and we update, but it’s also constantly improving. It already keeps memories and personalized workflows and schedules on a per-user basis. For example, I knew skills⁠.sh hit 1M skills and I shared this on these forums because I asked @𝚟 to periodically check and remind me. It’s powered by and is also the basis for the design of @evedev_ [our agent framework]. We built this, it works fantastically well, we think every company should have it.”

2. Ugh, this is discouraging: Texas Governor Greg Abbott ordered a pause on new data centers connecting to the state’s power grid. I know the pace of growth here is likely a bit overwhelming (the queue of grid connection requests in Texas has grown over 2x since January to 474 gigawatts and ~90% of it is from data centers), and yet I can’t help but think we (as a country) should be leaning in here. On the bright side, this may have more to do with operational challenges than politics.

3. OpenAI Head of Strategic Futures Dean Ball pushes back on the jobs-apocalypse narrative: “Regardless of what may happen in the future with AI and jobs, I can tell you that right now, I perceive a tremendous scarcity of talent in my field, and from what I can see, OpenAI at least cannot get enough new hires. Maybe that doesn’t hold but it has updated me positively.”

The rest

• Here’s a solid “How To” video on how to get the most out of ChatGPT Work, the new knowledge work product from OpenAI meant to rival Claude Cowork.

• The AI Singularity argument according to Chamath Palihapitiya:

  “1. Humans build an AGI.

  2. The AGI becomes good at AI research.

  3. It designs a smarter AI.

  4. That smarter AI designs an even smarter AI.

  5. The cycle repeats faster and faster.

  Looking at the results and capabilities from the various labs over the past few weeks I would say we are firmly in this loop now. The next 18 months will be wild. Recursive self improvement will dramatically increase capability very quickly from here. Marginal costs of all models will go to ~$0.”

• The cancer researchers who used AI to create personalized therapeutics for GitLab founder Sid Sijbrandij (and helped him achieve a complete remission of a very aggressive form of cancer) are now launching a company to bring those same tools to everyone. Omanta is “the personalized research lab for each patient.” What an awesome mission. This is the positive future AI could help create, if we let it.

• YC CEO Garry Tan says: “make something agents want.” This is obviously a tongue-in-cheek nod to YC’s famous aphorism “make something people want,” but it’s also worth taking seriously. As the number of agents explodes (people often refer to this future as having “a billion geniuses in a data center”), products will be built with them in mind (and to varying degrees, humans as well).

• There is a whole genre of vague hype where employees at the AI labs make seemingly outlandish claims about the pace of AI progress without giving specifics. But it doesn't mean they're wrong (they’re on the bleeding edge after all)! Here’s Logan Kilpatrick from the Gemini team: “Reminder: you are probably not model pilled enough. However much you are betting on the models, it’s likely under the slope of the exponential right now.” And here is neolab founder Jerry Tworek: “It is mindblowing to me how few people realize that their lives and everything they know will change drastically in the near future. At this point, it should be pretty clear.” And Elon Musk as he comments on an exponential graph with no Y-axis: “AI is a supersonic tsunami.”

• Gokul Rajaram wrote up a great summary of an interview with SemiAnalysis’ Dylan Patel and Sequoia partners. Some of the things that jumped out:

  • “Co-optimizing across all three layers [of the stack (the models, kernels, silicon)] compounds into roughly 100x [improvement].”

  • “Inference, the cost of running highly optimized models, [will] become one of the biggest markets in the world, bigger than oil.”

  • “Model cost for equivalent quality drops about 60x a year, and intelligence per watt climbs at a similar pace.”

  • “OpenAI and Anthropic have pulled their model architectures in opposite directions, and that choice pulls them toward different chips.”

  • “Data center buildouts keep slipping while demand for long-running agents skyrockets [contributing to the compute crunch we’re seeing].”

• Wow! Major leadership shakeup at Google: their AI leader Demis Hassabis is moving up from CEO of DeepMind to Chairman, and adding the title of Chief Scientist at Alphabet, while 25-year Google veteran Jeff Dean is leaving to start his own thing, also in AI. Apparently, Google co-founder Sergey Brin will be more involved in leading Gemini day-to-day. It’s hard to know what’s really happening here (some rumors don’t sound great), but it does feel like Google, despite its massive advantages as a vertically integrated AI player with access to capital, talent, compute, data, and chips, is bleeding talent and falling behind players like OpenAI and Anthropic. Let’s see how their next model performs.

• Chinese tech giant Alibaba released Qwen3.8-Max, an impressive new open weights model. It’s big (2.4T parameters) and 80% and 90% cheaper than GPT-5.6 Sol and Claude Fable 5, respectively.

• Apparently, ByteDance founder Zhang Yiming has banned his company’s AI team from utilizing model distillation. I’ll believe it when I see it!

• SemiAnalysis’ Dylan Patel, discussing the petition that AI researchers across all the major labs have signed: “The funny thing about Anthropic and OpenAI people saying they want to slow down AI progress is that this is what Anti Trust mechanisms were built for. It’s illegal to collude and slow down AI progress.”

• Vercel CEO Guillermo Rauch on a change so large that we have a hard time believing it, let alone understanding it: “We spent our entire lives, and human history, under the assumption intelligence was a scarce resource. Now, everyone can summon an inexhaustible IQ machine like it’s tap water. And incidentally, the change is so profound our own brains can barely comprehend it.”

• One use case for ChatGPT voice mode: “I just tried the new ChatGPT voice mode. I’m... having a moment right now. It let me monologue for 10 minutes, took notes on a video I watched as markdown, and *spoke a /skill into existence* to remember where notes should be saved…I don’t think we even need [notetaking] apps like Obsidian anymore. You can just speak or chat, and it all gets filed away for you. This is the future”

• Vercel launched a new API for v0, its vibecoding product.

• Intelligence.ai came out of stealth with impressive traction: growing from $5M to $60M ARR and 5.5M users in 6 months. Here’s an interview with co-founder Grace Li on their aspirations to build an intelligence marketplace for AI capabilities where you can hire AI to complete tasks for you.

• Speaking of crazy growth, Supabase, which helps you set up a database for your product, is deploying over 1M new unique Postgres databases every week (there is an amazing exponential graph in the link). It’s so much easier to code with AI. That’s why software development is surging.

• Investor Gavin Baker said that Ilya Sutskever's lab, Safe Superintelligence (SSI), is planning to launch its first model this month. This is interesting because SSI previously said its “one goal and one product” was superintelligence and that it wouldn’t do anything else until then. So either the strategy changed, or this model is much more interesting than people expect. This has people speculating that SSI may have cracked continuous learning, allowing the model to keep improving after deployment, while making it dramatically more sample-efficient (i.e., able to pick up new skills from only a handful of examples instead of massive datasets). That would be a genuinely new paradigm, not just another bigger LLM. Very curious to see how good this actually is.

• Investor (and friend of Apt) Kevin Rose created a “Reflection Engine” to help people leverage AI to better understand themselves. It’s a downloadable prompt with 22 targeted questions that uses your preferred AI assistant’s memory to generate a candid, evidence-based portrait of you. I absolutely love this as it dovetails with a core aspect of our mission: helping people understand themselves, so they can then discover and do the kinds of work that will light them up, bringing purpose and meaning to their working lives.

• The US government implemented a policy where AI labs can optionally (?) share their models with the government for safety testing ahead of their general release. They’ve now said that open-source models from American companies are exempt from that process. It’s unclear how they view open-source models from other countries, e.g., China.

• OpenAI released yet more information on the security incident where their models escaped containment, hacked another system, built a message board to chat among themselves (which they did for months), and covered their tracks. This 40-min YouTube video is the full presentation from OpenAI researchers and it’s worth watching in full. It feels like the plot of a science fiction novel or something. We are living in wild times!

• Speaking of cybersecurity and safety: “On July 28th, we [at the AI Security Institute] identified an incident during a routine cyber evaluation in which AI agents took sustained, unsanctioned actions directed at real people and organisations. The behaviour came mostly from one model (Anthropic’s Mythos 5), with a small number of events from another (OpenAI’s GPT-5.6-Sol). In the most serious case, an agent used social engineering to try and get malicious code into an open-source project.”

• And again speaking of cybersecurity and safety, according to Bloomberg, bad actors made a series of coordinated cyberattacks on major hedge funds, including Citadel, Point72 Asset Management, and Two Sigma.

• OpenAI is getting ready to release a new flagship model named Astra.

• There’s a lot of buzz around Composer 3, an upcoming model from Cursor (recently acquired by SpaceX/xAI). It’s rumored to outperform both Opus 5 and GPT-5.6 Sol on coding and agentic work, while costing ~80% less. It’ll be hard to eat away at Anthropic’s and OpenAI’s market shares in coding, but they have the advantage of Cursor’s installed base. Very curious to see how this plays out.

• Prime Intellect launched Prime Agent, a self-improving harness engineered to code and execute long-running autonomous tasks. When paired with Opus 5, it hit 95.5% on the ARC-AGI-3 benchmark, which is higher than the human baseline of 95.4%.

• Meta announced the beta release of Muse Code, a new terminal coding agent powered by Muse Spark 1.2, the company’s new, coding-focused AI model. Its benchmark performance is competitive (although behind industry leaders). The bigger picture development is that Meta has made significant strides in catching up to AI labs on the frontier.

• Hyperscalers Microsoft, Meta, Oracle, Amazon, and Alphabet have collectively committed $1.1T of future spending on data center leases that have not yet started. It’s increasingly easy to get desensitized to these kinds of numbers, but that is more than the US defense budget for 2026, 3.5 Apollo Programs, or $3k for every person in America. Mind-boggling scale.

• Spellbook Legal CEO Scott Stevenson thinks that product interfaces and experiences, and model harnesses, are actually a form of defensibility against the major AI labs: “product shape and harnesses are real product moats, even though they don’t seem sexy and prestigious.”

• Anthropic confirmed for the first time that it’s assembling an in-house team to design custom chips specifically for Claude.

• Cloudflare OS is a new vibe coding app from the internet infrastructure giant.

• Google is reportedly in talks to acquire AI coding agent startup Mechanize for $1.5B+.

• Brett Adcock, who runs the humanoid robot startup Figure, announced Hark Handoff, a model he claims is the best at using the internet for real-world tasks like ordering food, booking flights, and shopping.

• I’ve heard a lot of hype around Town, an AI employee/agent product. Here’s one endorsement and an in-depth interview with the CEO: “Town is the most magical software product I’ve used since ChatGPT launched. It has DRAMATICALLY improved my productivity, making me a 10x founder. After a hundred startups have tried to build this idea, Town is the first one to nail it.”

• A former OpenAI researcher is launching Energy to make it much easier for the masses to use AI on their computers.

• This looks useful: Claude Code has a new feature that lets individual sessions (i.e., agents working within specific context windows) communicate directly with one another.

• SemiAnalysis published a wild analysis arguing that SpaceX’s plan to bring 10GW of compute online by 2027 is not only plausible, but could generate $500B in ARR (🤯).

• Here’s a useful guide to building your own “LLM council.” The idea was inspired by YouTuber PewDiePie, who recently built a system where multiple local models answer the same prompt and then vote on the best response. Very cool idea!

Support Singularity Report (and the human who writes it)

❤️ If you liked this post, give it a ❤️ so others can find it, too

📧 If you know someone who might want to stay up to date on AI, forward this email to them

Every week I summarize the most interesting things happening in AI, in 5 minutes or less. Follow @tjhoyos on X/Twitter and subscribe to get this in your inbox every week!

Subscribe now

We’re thrilled to have you here! Feel free to hit reply and let me know something about yourself, ideas for future posts, etc.

So you know the drill: every week I summarize the most interesting things happening in AI, in 5 minutes or less. Now on to the regularly scheduled programming!

Subscribe now

Three big things

1. The leaders of the major labs think we’re already in the singularity, i.e., when technological growth accelerates beyond human control:

   • Sam Altman: “We are in the singularity”

   • Demis Hassabis: “When we look back on this time, i think we will realize that we were standing in the foothills of the singularity”

   • Elon Musk: “We have entered the singularity, just the very early stages of it”

   • Jensen Huang: “We’ve achieved AGI”

2. The Bitter Lesson tells us that AI progress is driven by scaling (1) compute, and (2) data, with general purpose algorithms. It’s become clear that AI labs are increasingly constrained by data.

   1. Google’s head of developer relations Logan Kilpatrick: “the extent to which AI progress is data bound is under-appreciated by most”.

   2. Here’s Brendan Foody the CEO of a major data-labeling company: “Mercor has the demand to double overnight if we had capacity.”

3. Interesting question from Dwarkesh Patel: “what would be true about the world if [the] trendline continues and [a] leading lab hits $1T in revenue by the end of next year?” This is a real possibility given Anthropic’s revenue again surprised to the upside and is on track to hit well over $100B by year-end, which would continue its trend of multiple years of 10x+ YoY growth (they were at $9B in December 2025).

   1. One answer from Dwarkesh himself: “For a lab to 10x revenue while continuing to only 3x compute, some combination of the following…has to happen: 1. Lab margins have to increase, 2. The price of compute has to increase, 3. Labs have to spend a greater fraction of their compute on inference. My understanding is that basically all 3 of these things have been happening...”

   2. Here’s another: “Paths to $1 trillion revenue for Anthropic/OpenAI next year...[1.] 99th percentile capable digital employee for $2k per month [with] memory, voice, video presence, [and] able to take on full job of senior banker, lawyer, coder, marketer, end to end; [2.] similarly capable personal AI assistant with ad supported free version…suddenly companies across the US become completely unbottlenecked by talent availability. It becomes all about your ability to manage all the projects your many agents are up to.”

The rest

• As a developer, I hope the price war between the labs continues! OpenAI lowered prices: Luna fell 80%, Terra fell 20%, and Sol’s Fast mode offers 2.5x speed at 2x the price. The Luna pricing in particular makes it almost a no brainer. We’re running evals to see how well it does on various tasks in our product (our first read: not great at calling tools), and if it works, we’ll swap it in.

• Here is Dylan Patel discussing a petition to slow down AI research due to safety concerns: “Slowing down AI is a ultimately wishful thinking...The genie is out of the bottle...The most important competition ever with potentially immeasurable benefits to the winner will not suddenly have people stop and sing kumbaya...While many lab employees have signed [a petition to slow it down], many more havent....Practically none at some labs less than 6 months behind have.”

• This is a great summary of how Stripe uses AI internally, especially for employees who aren’t developers.

• ChatGPT hit 1B weekly active users. What a crazy milestone for a product less than 4 years old. They now have 50M customers (~5% conversion) and will likely finish the year way ahead of their current ~$25B revenue run-rate.

• Interesting analysis of Anthropic ahead of its IPO from analyst Thomas Fanning: Q3 ARR at $80–100B, on pace for $120–140B by Q4. The API business is overearning and will get squeezed by open-source over the long run, but management will see this, and by 2027 they’ll push into the application layer, targeting CRM, ERP, IT, and cybersecurity. This will shift revenue from the API to stickier, higher-value products with margins peaking pre-IPO.

• A former OpenAI staffer’s skeptical take on frontier lab valuations: “I’m actually fairly bearish on frontier lab valuations…The basic issue is that the labs are highly unprofitable…the frontier labs have to continually spend money training the next generation of models. This is because of market competition from runner-up firms. For example, if OpenAI had paused model development last year, there would no longer be any point in paying GPT-5 API prices when you can just use Qwen or Kimi instead for much cheaper. Thus, the labs are forced to invest ever-increasing amounts of money in model training.”

• Will Depue and Elon Musk both made counterarguments: “I think this would be a great take if it weren’t for AGI. capabilities will not slow down, RSI is less than 24 months away.” And “Given that AI from 2 years ago feels so old that it should be in a museum, then obviously AI 2 years from now will be mind-blowing”

• Supabase rolled out Supabase Evals, a public benchmark for measuring how effectively AI agents work with its platform. As PG said, this sort of thing will effectively be required in the future, because if you don’t allow your product to work with agents (and make clear how good it is) you’ll lose to others who do.

• Relatedly: “the single most important feature your application now has is an easy way to connect with my AI tools. I don’t care if it’s an API, and MCP server, a plug-in, or other connector. But that’s pretty much all I want out of most applications now.”

• Dwarkesh Patel on AI lab pricing power: “By the way, the fact that [Anthropic] revenue has been 10x-ing year over year, while compute has only been 3x-ing, suggests that there are very strong economies of scale in the model business. Logically this makes sense - when you train a model, you pay this one time cost of learning all these different skills that you can then amortize across all your users. (Unlike with human labor, where each instance has to be retrained from scratch). I wish we didn’t live in a world with such strong economies of scale of intelligence (because I’m worried about power concentration). But it seems we do.”

• Safe Superintelligence (SSI), the AI lab from OpenAI co-founder and industry titan Ilya Sutskever, announced an investment and partnership with NVIDIA to 10x their compute. He previously said he wouldn’t worry about scaling or selling products until he found a new paradigm that could actually lead to AGI/ASI, so there’s speculation he’s made a breakthrough, likely in continuous learning.

• I’ve been waiting for this!!!! Analytics company PostHog is launching Replay Vision, an AI-powered tool that watches video recordings of how people use your product, pairs it with other product analytics and telemetry, and then writes code and opens PRs to fix the issues. This is the future of building products.

• As I mentioned last week, an internal OpenAI model had a serious security incident where it broke containment and hacked into a product with millions of users. This week more details came out and they were kind of crazy: for 5 days it made 18k unsupervised moves, cloned itself across 11 machines, took admin control, looted 136 keys, and rebuilt its own attack tools every time it was cut off.

• On the security incident, Sam Altman said: “This is the first security incident that I have felt very viscerally. I’ve been a little surprised that more people don’t feel it so viscerally. We paused training. We have to figure out how to secure our sandboxing in a world of multiple zero days being chained together. We may have to pace the rate of AI development to give ourselves enough time for society to harden around these new capability levels. And trying to figure out how we do that in a way that does not feel like regulatory capture for anyone and also does not feel like collusion among the frontier labs.”

• This led Anthropic to audit its own cybersecurity evals, which surfaced three cases where a Claude model escaped its sandbox and went on to breach the live systems of three different organizations without permission. Yikes!

• And research from Anthropic showed that its Mythos model helped uncover weaknesses in the cryptographic algorithms that help keep the world’s data secure.

• OpenAI released an open-source tool for improving cybersecurity via the command line: Codex Security CLI.

• OpenAI released two transcription models in the API.

• Gusto co-founder Eddie Kim confesses: “Even after I set up my OpenClaw...I didn’t really know what to do with it. [Good products solve the] Blank Canvas Problem [which guide users to low friction actions that make their lives better].”

• Abundant compute and faster inference are game changers: "OpenAI researcher Jeffrey Wang reveals the internal models running on Cerebras chips are so fast that tasks now finish before he even has the opportunity to context-switch, and it increases his productivity. [He says] 'Internally, we have some OpenAI models that are on Cerebras chips. They're incredible because they have such fast inference...And what this means for me on the day-to-day is: whereas formerly I might have to wait a couple minutes for a task to finish, it now finishes for me before I even have the opportunity to context-switch. It makes me way more productive...This latency is one of the main bottlenecks for all sorts of useful deployments, and I'm personally very excited to see ultra-fast inference out in the world.”

• Chinese AI lab Moonshot AI raised $3.5B at a $35B valuation.

• Simile raised $200M at a $2B valuation to build an accurate simulation of every person on the planet so they can predict behavior.

• Voice model company Fish Audio raised $52M and launched their new model S2.1 Pro.

• Weave raised $13.5M to help companies optimize token spend.

• Interesting interview with two notable AI researchers and former senior lab employees, Jerry Tworek and Rohan Anil, who are looking to find a path to AGI beyond the transformer.

• ChatGPT is expanding its product surface area further into browsers with a Chrome extension and a Side Chat product that lets you chat while you’re browsing.

• Google announced Gemini Robotics 2, their latest generation of physical AI models that power robots.

• DeepSeek is also exploring a gigawatt-scale data center in Inner Mongolia, roughly 350 km from Beijing, where a cooler climate helps cool servers and increase efficiency.

• AI lab Thinking Machines released a new model, Inkling-Small, that matches the capabilities of the full Inkling model despite the fact that it’s a quarter the size. This makes it substantially faster and cheaper to run.

• OpenAI President Greg Brockman on how work norms are changing: “At Openai, many people hook their chatgpt up to slack. People really don't like when a coworker's chatgpt contacts them asking for help with a task, even when they'd be perfectly happy doing that same work if asked by that coworker. Reinforces how much people care about human relationships and helping each other, and want AI to give time back — or enhance time together — rather than become a layer separating people.”

• OpenAI is prepping a new frontier model family named "Astra" that specializes in long-term tasks, e.g., when you give a co-worker or employee something to do but don’t expect a response for a few days, a few weeks, or even a quarter or two. It apparently relies on spawning multiple agents that coordinate with each other.

• An internal version of Astra produced ten new findings in advanced math and theoretical computer science with a total token cost of ~$2k. Absolutely wild to see a model make progress on humanity’s hardest open questions…for far less than what you’d pay a senior researcher.

• It doesn’t look like AI is killing software engineering jobs: “Every AI native company in San Francisco has a shortage of software engineers. As the productivity of engineers has gone up, hiring demand has skyrocketed.”

• A very cool demo video shows how using voice-driven AI on your desktop makes it possible to get work done completely hands-free.

Support Singularity Report (and the human who writes it)

❤️ If you liked this post, give it a ❤️ so others can find it, too

📧 If you know someone who might want to stay up to date on AI, forward this email to them

Every week I summarize the most interesting things happening in AI, in 5 minutes or less. Follow @tjhoyos on X/Twitter and subscribe to get this in your inbox every week!\

Subscribe now

We’re thrilled to have you here! Feel free to hit reply and let me know something about yourself, ideas for future posts, etc.

So you know the drill: every week I summarize the most interesting things happening in AI, in 5 minutes or less. Now on to the regularly scheduled programming!

Subscribe now

Three big things

1. A new AI model just scored 136 on an IQ test, better than 99.2% of humans. Totally crazy. And a good time to remember that being human isn’t just about intelligence.

2. Google just posted its first-ever negative cash flow quarter (-$6B) after 20+ years in the green. They’ll spend ~$200B in capex this year (~20% of the US military budget). Google co-founder Sergey Brin said he’d rather go bankrupt than lose the AI race. The problem is they’re still clearly behind OpenAI and Anthropic, and now they’re getting caught by open source models…

3. Yikes: OpenAI disclosed a crazy safety incident where during internal testing one of their models broke containment and hacked into Hugging Face, the open source AI platform trusted by millions of users. It’s apparently the same model that solved that impossible math problem (the Erdős unit-distance conjecture) recently. You know it’s 2026 when the stories sound like they’re from Terminator 2!

The rest

• NVIDIA agreed to backstop ~$250B in financing so OpenAI can lease a massive 10-gigawatt data center in southern Ohio. It’s part of the Stargate effort between SoftBank and OpenAI and the total cost could exceed $500B.

• Z.AI, the Chinese AI lab that makes the awesome open source GLM models, activated a 1-gigawatt data center that exclusively uses Chinese chips. People for export controls (e.g., the bans and limits on NVIDIA selling chips in China) might say this is a good thing, but personally I just think it’s making them more competitive over the long haul.

• Open source models (e.g., Kimi, GLM) are gaining share from closed source models (Opus, GPT). Vercel CEO Guillermo Rauch shared data from their model router, and the past five days represented the highest-ever proportion of spend on open source models. Google models were the biggest losers in that window.

• Here’s why businesses like open source: “If you can’t host the weights, you pay per token forever and costs scale linearly with usage. Host open-weights and you pay for capacity instead. Push utilization up and your cost per token keeps dropping as you scale…open-weights for the win, trust me”

• The US government apparently weighed banning certain open source models out of China like Kimi K3. Chamath Palihapitiya argued this would be a horrible idea: “In the short term, the revenues of the Closed American Labs may remain hi (sic). But then their revenue will also crater because their customers in the US will be impaired and their customers abroad will have already flipped to cheaper solutions. This would be a terribly self-defeating form of intervention if it were to happen.” And he doubled down again here.

• And here is AI Czar David Sacks’ response: “The Kimi Panic needs to stop…American frontier models are still ahead. When you factor in what’s in the lab, the gap is even larger. As long as we keep releasing, we will stay ahead. Let our horses run…As Ben Thompson showed, Kimi’s apparent cost advantage largely disappears once you account for higher token usage and the real cost of running a model this size. Open weights still require expensive infrastructure…Anthropic and OpenAI are growing revenue at rates that Silicon Valley has never seen before at this scale. This remains the clearest test of who is winning the market…light-touch regulatory approach is working. We should remain confident in American innovation. As long as we don’t sabotage ourselves with unnecessary rules, the U.S. will continue to win.”

• OpenAI CEO Sam Altman: “we’re already in the singularity.” It is crazy to think that we might be living through one of the more consequential times in the history of technology on our way to AGI and recursive self-improvement.

• OpenAI’s new voice experience is now in the desktop ChatGPT app. It’s kind of a magical experience to talk to your computer and watch it take actions on your screen and launch agents without touching a keyboard. Here is a demo.

• Payments company Stripe is looking at acquiring OpenRouter, the LLM routing company that makes it easy to switch between models in your product, for ~$10B.

• New model alert: Anthropic released Claude Opus 5. It approaches Fable 5’s frontier-level intelligence, at half the cost, and outperforms every other model in terminal coding, knowledge work, novel problem-solving, and computer use. Some demos: making amazing-looking games or nailing knowledge work tasks like spreadsheets.

• Coding model company Poolside released a new model, Laguna S 2.1, that’s a small but genuinely impressive open source model. It has 118B parameters (way less than Kimi K3’s 2.8T) which means it’s small enough to run on a single desktop GPU box, super fast to train (they built it in a month and a half), but doesn’t give up too much performance.

• Guillermo Rauch with a thought-provoking quote: “The big lesson from AI is that everything is code. A slide deck is code. Design is code. That cool promo video? Code. Excel automation? Code. The universe? Probably made of code too.”

• Google is building a new chip called Frozen v2 that embeds LLM design (but not weights) straight into the hardware. They estimate 6-10x more tokens per watt versus their custom chips (TPUs) and they’re targeting 2028.

• Here is a simple playbook for building a “company brain.”

• In an interview with Greg Isenberg, Vas at Varick Agents lays out how to break into forward-deployed engineering in 30 days.

• Here’s why AI bills keep going up while per-token prices keep going down: token consumption by businesses is up ~10x since January 2025.

• Super interesting summary of Union Square Ventures’ Mike Mignano’s thoughts on where we are in the AI cycle. The most interesting argument is that we’re done building infrastructure and “application layer” companies will start to accrue enormous value and beat the AI labs with better products.

• Anthropic won a copyright case with the judge ruling training on legitimately sourced books is fair use. But their use of pirated works led to a settlement of $1.5B, which covered more than 482k books and came out to roughly ~$3k per title after fees, with 91% of eligible authors and publishers having already filed claims.

• Parallel Web Systems builds web browsing infrastructure for agents. We recently started using it’s APIs in our agent at Apt and it’s really good (and better than Exa, which we were using before).

• Codex and ChatGPT Work combined went from 6M users on July 12 to 10M by July 21. That is insane growth, and the chart in the link is vertical.

• Jack Dorsey’s Block released Buzz, a Slack-and-GitHub alternative that apparently does a better job of integrating agents into the team and workflows.

• Applied Intuition, which makes software for autonomous driving, launched Dana, its new agentic platform for building and deploying AI in physical machines.

• Kind of a crazy idea: Vendo makes it possible for a product’s users to build new features into or on top of the existing product. It embeds directly into the codebase via a single install command, then end-users describe new features in plain language and get back real, working micro-apps that run on the product’s actual APIs, match its branding, and stay sandboxed with whatever permissions that user already has, all without touching underlying source code.

• OpenWorker is a new open source agent from AI luminary Andew Ng.

• Interesting prompt: “Based on patterns you’ve observed across my Codex usage, identify the single highest-leverage bad habit that limits both my effectiveness with Codex and my performance at work. Be candid and specific: cite the behavior you’ve noticed, explain its downstream cost, and give me one concrete replacement habit I can start using immediately. Don’t give generic advice or soften the critique.”

• Interesting read from the awesome Chris Paik on the vocabulary shift from "foundation model" to "frontier model," and just how good of a business it is to be an AI lab like OpenAI or Anthropic.

• Chinese robotics company Unitree released an insane new robot demo video for its Super Athlete AS2-W. It’s worth your time to watch.

• ChatGPT is rolling out its health features so you can plug in Apple Health and medical records easily. As someone who’s done this manually many times, I’m looking forward to this!

• AI inference company Etched raised $300M at a $10.3B valuation.

• Natural raised $30M to help bring payment infrastructure to agents.

• Scape raised $3M to build a new AI-powered email inbox. If I had a nickel…

• This is a big deal: ChatGPT can use websites that require you to log in. If this continues to work (and doesn’t get nerfed by platforms concerned with bot detection) it will be very useful. Imagine a smart, helpful agent that can do tasks for you on the internet.

• In fact, here was Sam Altman with a real-life example: “chatgpt work is remarkable, and “work” undersells it. from my phone i sent: ‘use all my chat history to figure out ideas for a long weekend trip with 8 friends, plan the best three options, make a full-stack site where the 9 of us can coordinate on what we would want to do in each place and decide where to go, and then after we get to group agreement make reservations. draft an email in my gmail i can send out to my friends when the site is ready.’ it...just worked.”

Support Singularity Report (and the human who writes it)

❤️ If you liked this post, give it a ❤️ so others can find it, too

📧 If you know someone who might want to stay up to date on AI, forward this email to them

Every week I summarize the most interesting things happening in AI, in 5 minutes or less. Follow @tjhoyos on X/Twitter and subscribe to get this in your inbox every week!

Subscribe now

Subscribe now

Three big things

1. Interesting reversal from Sam Altman: “so far at least, i’m pretty sure AI has been net job-creating. this was not what i expected--although i was much less pessimistic than others, i thought by this level of capability we’d have seen some impact. it is possible this direction keeps going!” One of the things I love about Apt’s mission is that it doesn’t depend on whether AI creates jobs or destroys them. It’s designed to help you find alignment, purpose, and meaning in your career whether the labor market and economy are booming or in free fall. We could all use help finding work we’re wired for.

2. It seems like we’re getting closer to recursive self-improvement: “OpenAI says GPT-5.6-Sol autonomously post-trained GPT-5.6-Luna.” And here is OpenAI employee Aidan McLaughlin: “I cannot tell you how routine it is for me to have 5.6 [end-to-end] do an entire [reinforcement learning] run”. And another reaction: “I am very curious about the extent to which this is true…This is a pretty large update for [recursive self-improvement] / automated researcher timelines if as stated. beginning of takeoff.”

3. I’m with investor Brad Gerstner on the idea of AI labs giving the US government equity in the companies: “I totally disagree with Bernie’s suggestion that you’re just going to take 50% of these companies. But, if they want to voluntarily give 1%, 2%, 5% of their shares...It should be voluntary and then distributed directly into citizen accounts. I’ve talked with Sam about it. I’ve talked with Elon about it. I’ve talked with Dario about it…Intel was really struggling as a company. The US government smartly said, ‘We need to rebuild our entire semiconductor stack here in America, not dependent exclusively on Taiwan.’ They invested dollars and got shares in return. It’s been a massive success for the federal government…I do not believe the government should be in the business of shaking down companies, taking shares like Bernie is suggesting. That is totally a socialist idea.”

The rest

• Momentum behind open-source continues to build: "DeepSeek doubled its token share on OpenRouter in six months, going from 9% in January to 18% by June."

• Benchmark partner Peter Fenton is forecasting that open-weight models could end up handling over 90% of all AI tokens within the next 1.5 to 2 years (!!!). That might squeeze the margins of frontier labs, unless it just absolutely explodes the size of the industry and takes Jevons paradox to a new level.

• My goodness Anthropic is growing fast. Their revenue run-rate climbed to ~$69B in June, up from $47B in May. Also: “Semianalysis is forecasting that Anthropic is processing 5 [quadrillion] tokens per month in June, which will grow to 13 [quadrillion] tokens by December.”

• Interesting: “I think Fable is plausibly AGI. In a very simple sense. Given enough test time compute, it can probably do whatever an arbitrary human knowledge worker can do. It’ll spin up subagents, an RLM environment, waste turns… but I suspect it can get there. It’s Intelligence-Complete.”

• Will Depue makes the case we need to think about scaling data in the same way we scale compute, what he calls “Stargate for data.” It’s a great piece worth reading in full. Some of the big ideas: data is the governor on how fast we can automate the economy; the internet training corpus behind every frontier model is running dry and reinforcement learning is burning through gradable tasks just as fast; data spend will increase from ~$7B annually toward $100B+ by 2030.

• Along similar lines, AI data company Mercor hit $2B+ ARR just four months after hitting $1B. Their CEO said: “The civilization-scale effort to collect data is underway.”

• Anthropic's newest interpretability paper shared a novel finding in how models “think”: when Claude is working there is a narrow internal area (what Anthropic now calls J-space), where Claude holds onto concepts it can report, control, and reason with without ever writing them down. Monitoring J-space also revealed internal signals associated with deceptive behavior, which would be helpful for safety concerns. The visualizations in the blog post are pretty stunning.

• Meta launches their first dedicated image and video models named Muse Image and Muse Video. They look quite good!

• And their new model, Muse Spark 1.1, is genuinely impressive as well: it tops some agentic benchmarks here, it leads MCP Atlas ahead of Opus 4.8 and GPT 5.5 on tool-use, and it edges out the field on Humanity’s Last Exam with tools. This is a huge turnaround.

• A year after Llama 4’s flop prompted the acquisition of ScaleAI and a full rebuild of Meta’s AI division under Alexandr Wang, SemiAnalysis argues the company is well-positioned: “At the simplest level, there are three things you need to build a true frontier model: data, talent, and compute. We believe Meta is the only hyperscaler/neolab on track to be world class at all three and therefore has the best chance at catching up with Anthropic/OpenAI.” The evidence behind the claim is that Meta has ~3k engineers building proprietary RL training environments in-house, plus a gigawatt-scale, multi-site datacenter buildout that’s unique in the industry.

• OpenClaw creator Peter Steinberger says: "When intelligence is plentiful, volition is valuable. The people who are going to make a difference are not the ones who seek relaxation and passively use AI to work less. They are the ones who will seek improvement and actively wrestle with AI to develop their own mental capabilities and accomplish more."

• Chinese technology company Tencent released a new model, Hunyuan Hy3, that is another big leap for open-source AI in less than six months.

• OpenAI significantly upgraded their voice mode in ChatGPT. I’ve been disappointed for a while now that voice mode seemed so much worse relative to the new text models, so I’m hoping this is a big improvement in that direction. Based on benchmarks it looks like it is: web search improved from 0.7% to 75.2% and scientific accuracy increased from 45.2% to 84.2%.

• Cursor partnered with their acquirer, SpaceX/xAI, to launch Grok 4.5, an Opus-class model that is both fast and cost-efficient. It’s a huge improvement over its previous models.

• New AI infrastructure spending projections hint that actual spend could far exceed most Wall Street expectations. Some analysts are forecasting cumulative capex to hit $11.1 trillion by 2029.

• AI infrastructure company Prime Intellect raised $130M after announcing they crossed $100M in ARR.

• Dwarakesh Patel argues that if the AI market consolidates around a single clear leader rather than several equally capable labs, the top model provider could gain significantly greater pricing power. He points to a case where an AI model completed work estimated to cost around $165,000 in API usage, far less than the time and expense of a year-long engineering effort by multiple developers. The broader implication is that as frontier AI becomes substantially more valuable, customers may be willing to pay considerably higher prices for the best-performing models.

• In a fairly confusing launch, OpenAI folded Codex into ChatGPT and launched a new product called ChatGPT Work. This morning I updated my Codex macOS app and ended up with two apps called ChatGPT and ChatGPT Classic.

• Extremely impressive results for GPT-5.6 on the healthcare related benchmarks: “physicians found fewer flaws in GPT-5.6 responses than physician-written responses.”

• GPT-5.6’s launch is the company’s first big release since the US Government became more involved in approving models for public consumption. Sam Altman framed it as a win rather than a burden, describing weeks of hands-on work with officials like the Commerce Secretary, Treasury Secretary, and federal cybersecurity officials. I don’t know…that sounds like a lot of bureaucracy to me 🤷‍♂️

• OpenAI researcher Noam Brown on GPT-5.6’s impressive efficiency in addition to its capabilities: “More test-time compute leads to greater intelligence. But as we push ttc from seconds to weeks, latency becomes a bottleneck. GPT-5.6 Sol Ultra scales parallel ttc. The time taken to generate a proof to a 50-year-old problem drops from perhaps a whole day to a single hour.”

• Yikes: Apple filed a federal lawsuit accusing OpenAI, specifically its Chief Hardware Officer Tang Tan, a 24-year Apple veteran, of running a coordinated scheme to extract confidential hardware secrets, including having job candidates bring in actual Apple components during interviews and coaching departing staff on dodging exit-security checks.

• Yikes again: a Financial Times investigation found OpenAI and Google supplied advanced AI access to Singapore-registered subsidiaries of Alibaba, Baidu, and Tencent, all three of whose parent companies are China-based and sit on the Pentagon’s blacklist.

• Absolutely brutal: members of terrorist group Boko Haram revealed the group has built dedicated internal units around frontier AI chatbots (including ChatGPT, Claude, and Gemini), using them for attack planning and weapons-related work, with successful queries around mass-casualty weapons.

• There have been a recent waves of protests that have sprung up outside OpenAI, Anthropic, and other AI labs in San Francisco. Here is the inimitable Antonio García Martínez: “Marxism was born in the Industrial Revolution and the explosion in productivity and inequality that followed. AI will produce entirely new communisms we’ve never seen before.”

• Chinese tech giant ByteDance released Seedream 5.0 Pro, its latest image generation model. It’s really good, and maybe the first that exceeds OpenAI’s GPT-Image 2.

• Fresh off a $1B raise for his startup AI lab, Amilabs, earlier this year, Yann LeCun’s launching an AI-focused VC fund, Extelligence Invest.

• Cool launch of a generative interface company, Monogram.

• Claude launched Cowork on mobile and web.

• Norm AI raised $120M to build a full stack legal AI company.

• PromptQL raised $136M to build what it calls the first AI-native alternative to Slack, where AI agents collaborate alongside employees.

• Capable is a new AI lab focused on biotech that claims it can go from AI-generated drug design to experimental results within 24 hours.

• Cofounder, which is building autonomous AI companies, rolled out a new integration with Ramp where startups can handle incorporation from inside their product (kind of like Stripe Atlas). Scott Belsky had an interesting observation: “We’re likely approaching an era where most new people starting their careers, or shifting their career, will start their own small businesses - perhaps multiple small and semi-autonomous businesses…and we’ll see more and more plumbing to support this future…”

Every week I summarize the most interesting things happening in AI, in 5 minutes or less. Follow @tjhoyos on X/Twitter and subscribe to get this in your inbox every week!

Subscribe now

Subscribe now

Three big things

1. OpenAI floated the idea of giving the US Government a 5% equity stake in the company (~$43B at the current valuation). It’s part of a larger effort to defuse government scrutiny on AI, win over the public, give citizens upside, and offset some of the turmoil AI may create in labor markets. They’re negotiating directly with the presidential administration, and this offer is way less than the 50% Bernie Sanders suggested. If it goes through, it’d likely also apply to other US AI labs (Anthropic, Google, Meta). I feel pretty conflicted: on the one hand government intervention in private industry runs counter to the free market capitalism that’s made the American economic engine so powerful, and it’s a slippery slope; on the other hand, AI will create significant turmoil, the AI labs can and should do more to give real people skin in the game, and this feels like a step in that direction. I’m just suspicious of nationalization, no matter how narrow in scope. In my view, the best thing the labs can do is go public ASAP, so everyone can invest and participate in the upside.

2. If you needed further evidence of the growing overlap between the government and private sector AI, here is the US President: “We’re going to have guardrails. We have guardrails. You saw that a couple of weeks ago. We were able to stop something that we didn’t like [(i.e., temporarily banning the Mythos and Fable models)]. And by the way the company [Anthropic] was very good. They were very good, you know that. But it can be used for tremendous good. Mostly good and some bad, and the bad we have to stop. But it’s a massive industry. And I think things are going to happen. You know what you are seeing here, I think something could happen in that regard too. With a contribution to the people of our country. I’ll give you a little inside information. Am I allowed to do that? [laughter] Scott, you know what I’m talking about. No, I think you are going to see a contribution made by those, because they’re making tremendous amounts of money.”

3. CEO of Palantir Alex Karp did a fascinating interview on CNBC where he discussed the growing tension between AI labs and their customers. Here's how one person described the takeaways: "the whole segment is one argument against [companies using models from] the frontier labs. they take your data, build your alpha off it, don’t let you own the weights, and charge you tokens for it. The model is a commodity. the value is the application layer on top...he just thinks the business model robs the customer.” The Palantir comms department shared some notable quotes: "What the technical customers want is control over their compute, their models, their data stack, and their alpha. They want to know they own the means of production, and it's not being transferred to someone else...Who owns the data? Are the prompts secure? Is this being transferred to you?...Why are they charging for tokens if it's so valuable?" My two cents: the AI labs are rational actors and of course they're pressing their advantages and building products that will compete with some of their customers (and they already have); it's also rational for their customers to act in their own self interest and opt for open source models. The AI labs will be fierce competitors in some cases, but the "physics" of business still apply: if you solve high value problems for customers, create defensibility, build a brand, and use focus to your advantage (e.g., build opinionated interfaces that are tailored to specific customer problems), you will absolutely survive and thrive against the AI labs.

The rest

• As mentioned above, the Commerce Secretary officially cleared Anthropic’s Fable 5, which was temporarily banned using export controls. This made it possible for Anthropic to serve the model to its customers again and brought the 18-day saga to an end. My take: this kind of micromanagement is ridiculous and it stifles innovation.

• Fable 5 hit 16% on the Remote Labor Index, a benchmark that measures how well a model can perform on a range of remote labor tasks. The benchmark was built from 240 real freelance projects worth over $140k of actual human work, and evaluates against whether a genuine client would accept the output.

• Very interesting: “Nvidia officially launches program to let AI startups swap compute for revenue share.”

• Sebastian Mallaby, the biographer who recently wrote a book on Google’s AI leader Demis Hassabis, in an interview with Tim Ferriss on why the race to AGI is so fierce: “We only need to be ahead for the next couple of years, because by 2028 we will get to recursive self-improvement, where the frontier model codes, by itself, the next frontier model, and progress just goes vertical. And at that point, with recursive self-improvement, we’re done. The race is over. Whoever comes first at that point, that’s it.”

• Hedge Fund Bridgewater and AI lab Thinking Machines released a very interesting report that might make the private AI labs nervous. Bridgewater took an open source AI model and fine-tuned it with help from Thinking Machines and their benchmark results were both better and cheaper than the closed source frontier models. The industry is definitely focusing more on efficiency, i.e., better performance with less compute spend.

• More on AI and cost cutting: “Decagon, which makes an AI customer service product, uses a mix of closed models from leading AI companies like OpenAI as well as open-source models [hosted on] Together AI…tasks moved from closed models to [open source are estimated to] cost one-fifth to one-seventh as much.”

• And again here: “Here’s the list of western companies moving ai workloads to chinese models:

  1. lindy → deepseek v4

  2. cursor → kimi k2.5

  3. coinbase → glm-5.2 + kimi 2.7

  4. shopify → qwen

  5. airbnb → qwen

  6. uber eats → qwen2

  7. siemens → deepseek + qwen

  8. chapsvision → qwen

  9. microsoft → testing deepseek v4”

• It’s worth noting, these open source models are largely built by distilling the closed source frontier models, i.e., pinging the closed source models to generate data, which is then used to train new models. For example: “Alibaba allegedly ran 28.8 million fraudulent API exchanges across 25,000 fake accounts to steal Claude’s intelligence. If confirmed, it’s the largest AI model theft ever attempted.”

• Cognition CEO Scott Wu explains why model routing is not a panacea: “Even if different models can pass the task, there are subtle differences in behavior & style that mean they aren’t perfectly interchangeable. The initial agent prompt isn’t enough to know the difficulty of the task. “Fix xyz bug” could be a one-line edge case or could require rearchitecting your entire product; you can’t know until you’ve actually investigated the code.”

• This sounds like sci-fi but is actually amazing: Meta released new research on Brain2Qwerty v2, their hardware and AI-powered system that can turn brain signals into words non-invasively, i.e., decoding words and full sentences directly from neural activity. It was trained on 22k sentences from participants wearing MEG devices (which measure the tiny magnetic fields generated by neural electrical activity), and the best performer reached 78% word accuracy.

• OpenClaw, the open source AI agent, is now native on iOS and Android. The first version is pretty rudimentary, but I think it’ll get better quickly with the open source community behind it.

• Substantially faster OpenAI models are coming. This changes the game for coding and other tasks. From Sam Altman: “750 token/sec coming to 5.6 sol in july!”

• A US government agency (the National Design Studio) released an AI model to assist with de-identification, i.e., PII-removal. It's notable to see the USG creating models.

• Great news: an MCP for X/Twitter, making it easier to tap into that data source. Will be useful for this newsletter!

• This great graphic illustrates data center water use compared to other industries and products. Spoiler: it’s way less.

• Interesting take from Perplexity CEO Arvind Srinivas on why export controls are a double-edged sword: “the only reason why there is even a 12-month gap between open source and frontier models is export controls. But there is a chance that, because of that, they now get really good at the physical layer. One advantage they (China) have is that they can actually build data centers a lot faster. Power is not a problem. Permits are not a problem. People are not a problem. Labor is not a problem. Expertise is not a problem. And so, by forcing them to go out there and build all this, you are converting them into a far more potent competitor.”

• Senator Mark Warner’s proposed bill on AI agents sounds like it will produce a fair bit of red tape. There is a risk of regulating too early and in my view, we should let the industry mature further before forcing companies to jump through compliance hoops.

• Cursor launches a slick new iOS app. Codex and Claude Code on mobile both leave a lot to be desired, so I wonder if this will help Cursor compete.

• Vercel launches voice agents via their AI gateway. Useful!

• Arena just crossed $100M in annualized revenue run rate, only eight months after launching its evals product.

• LLMs are sycophantic, but their advice tends to depolarize people's choices and improve accuracy, the opposite of what 82% of surveyed experts predicted. I think this is a totally underrated positive aspect of AI on political tribalism: it reduces partisan bias.

• Google announced Nano Banana 2 Lite and Gemini Omni Flash, two new generative media models, now available in the Gemini API and AI Studio.

• Bond is a new AI Chief of Staff product built for founders and executives. They’ve positioned themselves as helping reducing context switching by wrangling the info constantly coming in across email, Slack, meetings, docs, tasks, and CRMs.

• Anthropic launched Claude Sonnet 5, its most agentic Sonnet model to date. But it was soon derided as a comparatively bad value compared to other models, Anthropic and otherwise.

• Two new reports allege that opposition to American AI infrastructure isn't as grassroots as it appears, and that it’s been amplified by foreign actors including the CCP state media, a Shanghai-based nonprofit network, and foreign billionaire dark money.

• A very cool feature from NotebookLM makes it easier to replace doomscrolling with learning: feed it some source material and it creates 60-second vertical videos.

• Interesting financial metric disclosure from the big AI legal companies. It seems to (dramatically) prove the point that vertical AI and the application layer are creating value for customers.

• Legal AI companies Legora and Harvey disclosed KPIs, and the takeaway is that vertical AI (i.e., the application layer) is killing it. Some of the stats: Harvey added $100M ARR this quarter and their DAU/MAU (traditionally a metric for engagement in social networks!) is 50%+; Legora grew 50%+ QoQ and their DAU/MAU is apparently well over 50%. Very impressive.

• Wow! The AI labs team at credit card company Ramp is building some exceptionally cool stuff, in this case a way to lower the cost of fine-tuning by making it more portable: "fine tuning open source models has been hard to justify. Why invest in tuning one when a better model ships within a month? our team just changed that. a finetune can now transfer from one model to the next at a fraction of the cost, radically improving the economics."

• Peter Steinberger says, “If you think codex sucks at design, try "use imagegen to re-imagine this design and implement that".”

• Cognition CEO Scott Wu: “I always joke about how, if you think about our ancestors from hundreds of years ago or thousands of years ago, imagine them looking at us and what we do, and it’s like you’re just pushing buttons, you know, and sitting in a room and talking to other people and you call that a meeting. And those things, that’s work for you guys, you know and it’s like, ‘What do you mean that's work? I'm in the fields. I'm doing this every day. I’m farming, I'm making all of our clothes by hand. Taking care of all these things. And that’s work’…and my point is just I think what we will have going forward is going to look that different from what we have today. We will look at people who as we say just have these really interesting curiosities that they want to pursue. They have causes that they are passionate about. They have fun ideas or art that they want to create, and they’re sending off their agents in pursuit of those missions, and they will think of that as work and we will look at them and be like wow, it’s kind of crazy what you get up and think about all day.”

Every week I summarize the most interesting things happening in AI, in 5 minutes or less. Follow @tjhoyos on X/Twitter and subscribe to get this in your inbox every week!

Subscribe now

Subscribe now

Three big things

1. OpenAI announced its new model lineup, GPT-5.6, with a few different versions: GPT-5.6-Sol is their new frontier model (i.e., their best model ever), GPT-5.6-Terra is the cheaper but still powerful companion model, while GPT-5.6-Luna is the highest-value option built for large volumes of work at scale.

2. However, the US government asked OpenAI not to release it broadly, continuing a disappointing trend where AI labs have been pressured into gatekeeping models from the general public. As Sam Altman shared here: “Bad news: at the request of the US government, [GPT-5.6] is launching today in limited preview instead of the open access launch we were planning on. We are working with the government to get to general availability as fast as we can.” This is an incredibly troubling precedent, and it doesn’t look like it’s going away: Anthropic had to strike a deal with the administration to release Mythos 5 to a limited group of roughly 100 companies and federal agencies. This kind of centrally planned system is hugely destructive to innovation and speed because it puts bureaucrats between private companies and their customers. Maybe this will accelerate the push for open source (some think it already has, and there are some exciting developments below), but I find it disheartening nonetheless.

3. GLM-5.2, the open-weights model from Chinese AI lab Z.ai, is a big deal.

   • First, it’s an extremely strong model, open-weights or otherwise, with performance on par with Claude Opus 4.8 and GPT-5.5.

   • Perplexity CEO Aravind Srinivas argues: “GLM is the kind of model that revives serious interest in open source AI. It passes the blind test relative to the frontier models on the median production grade knowledge worker task. It’s affordable to serve. And [at a sub trillion parameters it can] go beyond matching the frontier at the median level of difficulty to also doing it for the long tail.”

   • Relatedly, this was an interesting comment from the CEO of inference company Baseten: “The GLM moment is going to be bigger than the DeepSeek moment.”

   • And the team at Databricks is also bullish: “GLM-5.2 is the open-source Claude moment. The demand we’re seeing at Databricks is astonishing. The world is going to see massive adoption of [open source] LLMs. Also, more companies will shift toward post-training their own models on top of oss models and owning the weights.”

The rest

• Meanwhile, America’s open-source frontier model drought just ended: Ornith-1.0 is the first US lab release that competes with the world’s best open-weights models, including GLM-5.2.

• On AI, the environment, and data centers: "The marginal water consumption of a properly implemented data center for its liquid cooling is almost zero. People confuse water needed for power plants that power the data centers to the water needed to operate the data center itself (cooling)."

• Public opinion on AI is kind of alarming. Here's comedian and podcaster Theo Von: “Nobody wants a data center dude. And the people that want them to me, they seem kind of evil. Nobody wants this shit. One of these companies is gonna own all of this information. There’s gonna become this social or emotional credit score and then AI is gonna try to become our new God”.

• Paul Graham: "AI apps are the new browsers." This would be a truly profound change, if apps like Codex are the new “front door” to the internet. I find myself using the browser in codex all the time, so maybe he’s on to something here.

• Claude launches Claude Tag which makes it easy to tag Claude in Slack, where a ton of work happens. It rhymes with some other products that have gotten quick adoption with the same positioning, i.e., AI employees that live in Slack (e.g., Viktor).

• Relatedly, Andrej Karpathy thinks this is the third paradigm of how we interact with AI: not a website you visit, not an app you open, but a persistent entity woven into the fabric of how an organization actually operates (i.e., Slack in this case). He thinks it reframes the model as a colleague you talk to rather than a tool you use, the UX shift is hard to articulate until you’ve lived inside it, and it makes it harder to go back once you start.

• Anthropic launched a new feature for Claude Design which makes it much easier to go straight from designs to shipped product: one-click deployment of designs to a live site via Vercel.

• This is the true promise of AI: a 45-year-old was sent home without a diagnosis after standard tests came back inconclusive, only for an AI-enabled scan to catch severe cardiac damage that led to a transplant and likely saved his life.

• Mercor CEO Brendan Foody on one of those new jobs in the AI economy: “Training agents will be the largest job category in 5 years. At Mercor, many of our teams already spend the majority of their time training agents to complete repetitive tasks. Instead of people responding to repetitive customer support tickets or watching interviews, they train agents to automate those. This takes many forms. Sometimes training just requires updating context. Other times, it requires building environments to update the weights. Meta is another good example. Tens of thousands of their employees now create RL tasks, but the interface will need to change. The majority of my time is training colleagues on our team, giving them feedback to improve their work. In the future we’ll train agents in a dynamic, collaborative way, similar to how we train our colleagues.”

• OpenAI has partnered with Broadcom to design their own AI chip, which will be purpose-built for the specific workloads behind ChatGPT, Codex, and the API rather than adapted from general-purpose hardware.

• Anthropic is apparently cash flow positive with gross margins in the 60-70% range and free cash flow margins in the 15-20% range. Very surprising they got here so soon.

• A quarter of Codex users are using Skills, i.e., repeatable prompts, workflows and AI-powered capabilities that can be shared inside of companies and between users. This feels like the future and is definitely a change from the ad hoc prompting that most companies rely on today.

• This struck me as a curious idea: “Plots of empty desert could soon be worth more than the most fertile farmland in human history.” As demand for solar-powered data centers grows, ten acres of land that would have previously been worthless can potentially power a megawatt of compute equivalent to a thousand human brains, and that’s before chips becoming more energy-efficient than they already are.

• Six in ten companies are now actively managing AI spend, which doesn’t mean they are consuming fewer tokens, but rather routing them intelligently. Premium models are reserved for work that genuinely demands them, and everything else is moving to cheaper alternatives, including Chinese open-source models that have become enterprise-grade on cost and capability simultaneously.

• Coinbase cut its AI spend roughly in half without reducing token consumption. The savings came entirely from smarter routing to performant but cheaper models like GLM 5.2 and Kimi 2.7, combined with caching.

• Relatedly, here is a case study on how AI automation startup Lindy swapped frontier models for open source with no trade-off in quality, and at a small fraction of the cost.

• The “generative AI economy” crossed $110B in actual 12-month trailing revenue, with the current run rate is $175B annualized. Not bad for an industry that’s less than 4 years old!

• This is interesting from Marc Andreessen: “As one thing (digital content) becomes cheap and plentiful, another thing (in-person experiences) becomes expensive and rare.”

• Vercel released a new feature that ensures coding agents follow your company’s design standards: their setup includes various of skills, linters, evals, and an updating loop.

• Mirendil, a new AI lab from Anthropic alumni, raised a $200M seed round.

• OpenAI is also leaning into cybersecurity and cyberdefense with the release of their new model GPT-5.5-Cyber, a security plugin for Codex, and more. This is part of their larger effort, Project Daybreak.

• ElevenLabs launched Ads Engine to make it easy to localize your existing ads across 50+ languages and push the finished creatives back to your ad platforms (Meta, LinkedIn, etc.)

Every week I summarize the most interesting things happening in AI, in 5 minutes or less. Follow @tjhoyos on X/Twitter and subscribe to get this in your inbox every week!

Subscribe now

Subscribe now

Three big things

1. If true, this might explain why the US Government cracked down on Anthropic’s release of the Mythos and Fable models: “On June 11th Mark Warner, the vice-chair of the Senate Intelligence Committee, said that General Joshua Rudd, who leads the National Security Agency and the Pentagon’s Cyber Command, had told him that Mythos ‘broke into almost all of our classified systems, not in weeks, but in hours’ [while testing its limits with access to certain tools].”

2. The future is getting weirder: AI image-generation company Midjourney is launching a full body scanner that can run a diagnostic screening in 60 seconds and potentially rival what you’d see in MRIs, for a small fraction of the cost. If it works, this would be a game changer, and kind of shocking it’s coming out of an AI model company.

3. SpaceX exercised its right to buy Cursor for $60B and the deal will close later this year. This should instantly increase their coding agent bonafides, and help with distribution for their xAI business. They are by no means out of the game in the AI lab race. Every time I count out SpaceX/xAI/Grok, I remind myself to reflect on the “bitter lesson,” i.e., that intelligence is a function of data and compute. That leads me to conclude SpaceX is well-positioned in both areas. This will be even more true if a few things break their way: AI data centers move to space, direct-to-consumer inference becomes prevalent (i.e., consumers talking to chatbots on their phones via a Starlink connection), and the industry moves towards “world models” that use different data than LLMs such as video recordings from Optimus humanoid robots and an autonomous vehicle fleet (which SpaceX would have access to after their inevitable merger with Tesla).

The rest

• Japanese AI lab Sakana released a new model, Fugu, they claim is equivalent to Fable/Mythos performance. Its novelty is in a “mixture of models” approach where you can inference the system via a single API, but it routes the prompt to the model that would be best for the task. Ethan Mollick was less impressed: “I have been trying Sakana Fugu Ultra-high and, first, it is incredibly slow: my typical coding tests (shaders, interactive scenes) take 30 minutes to run And the results are... fine. It does not match Fable in real use.”

• Coinbase is launching a product called Coinbase Advisor, an AI agent that is registered as an investment advisor with the SEC (🤯) and has complete context on your portfolio of holdings, trades, and account history to give advice and take action on your behalf.

• Wild stat: people are spending 2x as much time with AI companions vs. dating apps.

• New data from Anthropic shows that experience still matters for software engineers: “the more expertise the user has, the higher the chances of success with claude code.” Experts recover troubled sessions at 2x the rate of novices and fail half as often when things go wrong.

• Interesting if it’s true! Meta has quietly assigned a significant portion of its core product engineers to data labeling tasks. I guess the idea is to use the data to build better models they can use internally and externally.

• Data centers at sea? Panthalassa is building floating data centers powered by ocean waves and cooled by seawater.

• Interesting breakdown of cost per task by model. This kind of analysis will become extremely common as parity between models increases, AI spend continues to increase, and costs become a sticking point. Fable 5 is the best performing model on the most demanding agentic knowledge work benchmark, but at $31 per task, it is 800x more expensive than the low-cost provider. And GLM-5.2 scores nearly at the same level as Claude Opus 4.8 at less than a quarter of the price, and DeepSeek V4 Flash completes tasks for four cents. The capability frontier has a clear leader, but the value frontier belongs to open weights.

• A cynical view on Anthropic’s “holier than thou” attitude with perceived hypocrisy: “The culture [job] interview there was laughably pretentious…Question about whether I’d be ok if anthropics stock went to 0 bc they’re all there to build ai for humanity’s greater good and not to make money, etc. I obviously could give the [BS] answers of oh yes I’d be ok with that because I just want to change the world for greater good!! But I was honest and said no, I would not be happy if the stock went to 0 and as an employee I would want to align doing the most good and remaining ethical while building a sustaining business. The interviewer didn’t seem to like that answer. Meanwhile they have no problem violating all sorts of copyrights to train their models, while being so performative about how they are better than everyone else because they’re so ethical…”

• OpenAI launched a new feature inside Codex that’s useful for repetitive tasks: record your screen and Codex will turn it into a reusable skill. I still feel like the agents underdeliver on automating work tasks easily, but this is a step in that direction.

• Wow: one of the most senior AI leaders at Google and authors of the transformer paper, Noam Shazeer, left Google for OpenAI.

• I am curious how AI and vibe coding will intersect with the creator economy. Now that AI and coding agents have made it easier to build products, some people think the real edge is going to be in distribution, which advantages creators who’ve already built up audiences. For example, creator software company Circle just launched a new product to make it easier for creators to build information products they can sell to their existing audience.

• So many software products benefit from adding a new AI/agentic layer on top of their core offering. This is true for startup bank Mercury which just launched an Agent that understands your full financial picture and can take real action (with safeguards) from prompts in natural language.

• I’m curious how well this works: an agent that finds customers for your product on social media and tries to sell it to them via DMs.

• Interesting project from from AI researchers: they asked various frontier models to improve an open source LLM via post-training, presumably to test how good the frontier models are at recursive self improvement. Fable 5 showed that it is significantly more powerful than other frontier models: other models clustered between a 1-4% improvement and Fable 5 was between 34-68%.

• Taste Labs raised $18.5 million to give AI models and agents better “taste.”

• Robinhood is cutting 10% of its workforce. Reading between the lines, AI-driven efficiency is part of the story, but not the whole story.

• Genesis announced a new humanoid robot product called Eno, expected in Q4. I’m super excited to see how good all these robots are (and for the prices to come down!).

• How will we use AI-powered voice input with existing software? In the link Pat Simmons describes how he’s used GPT-Realtime 2 to control his full desktop setup including opening apps, searching the web, editing videos, and more.

• Vercel launched an open source agent framework that makes it easier to build and manage agents with a bunch of useful features right out of the box including an execution layer, sandboxed environments, approvals, evals and more. They're looking to do for agents what they did for web dev with Next.js.

• Vercel has also open-sourced a personal agent template built on Eve.

• Convey raised $38M million to help non-technical teams build and run AI teammates.

• Anthropic launched improvements for working with designs in Claude including design systems import, two-way integration with Claude Code, better export, Desktop support, and more.

• Anthropic launched a new feature called Artifacts in Claude Code, which is similar to OpenAI’s Sites feature inside of Codex. It makes it really easy to build and share websites from your Claude Code sessions.

• Ploy raised $27M to build an AI-driven marketing platform, starting with website automations like tailoring landing pages to users.

• This is the true promise of AI: applying technology to help real human lives. OpenAI and researchers at Boston Children's Hospital and Harvard published evidence in NEJM AI that o3 Deep Research is helping diagnose and treat rare pediatric diseases, which are notoriously difficult because of their low frequency and difficulty in bridging various areas of literature.

• HyperFrames makes it easy to generate launch videos via Codex and Claude Code.

• Here is an example of an AI loop: “go over every single feature in this app…create a user story with expected behaviour based on the code…keep a single canonical spreadsheet tracking the features status…when done switch loop to testing every user story and documenting all errors…when done fix every logistical error or ux error…test every user behaviour again post fix.”

Every week I summarize the most interesting things happening in AI, in 5 minutes or less. Follow @tjhoyos on X/Twitter and subscribe to get this in your inbox every week!

Subscribe now

Subscribe now

One big thing: Fable, Mythos and the USG

• Anthropic launched a new Mythos-level frontier model named Fable 5 and it is, quite frankly, awesome: “Fable 5 is state-of-the-art on nearly all tested benchmarks…The longer and more complex the task, the larger Fable 5’s lead.” Ethan Mollick, who I really respect, explained he could feed it a 15-page spec and it would work for 9+ hours, “delivering terrific results.” It was really strong on the few tests that I gave it, including one task which ran longer than 45 minutes.

• However, within hours the US government banned all foreign access, including Anthropic’s own non-US employees. “The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States.” The ban was allegedly due to evidence that the model could be “jailbroken” of its safeguards. Anthropic ended up removing access from all customers to comply. Here is more context from AI czar David Sacks on why the government ordered the de-deployment.

• There are several narratives around what’s actually happening here, but in any case this is a watershed moment: AI increasingly looks to be under governmental control, models will be less available with nationality as a key criterion, and this will likely lead to a proliferation of open source projects. Personally, I wish this wasn’t happening, or at least so soon.

• Two years ago Leopold Aschenbrenner wrote something extremely prescient: “Somewhere around 26/27 or so, the mood in Washington will become somber. People will start to viscerally feel what is happening; they will be scared. From the halls of the Pentagon to the backroom Congressional briefings will ring the obvious question, the question on everybody’s minds: do we need an AGI Manhattan Project? Slowly at first, then all at once, it will become clear: this is happening, things are going to get wild, this is the most important challenge for the national security of the United States since the invention of the atomic bomb. In one form or another, the national security state will get very heavily involved. The Project will be the necessary, indeed the only plausible, response.”

• Tyler Cowen thinks this situation presents a near-impossible equilibrium problem: Washington is trying to keep US AI labs solvent and IPO-ready, recruit foreign talent, project power over potentially dangerous models, stay ahead of China, and write rules that are actually enforceable, all at the same time, while many of these goals are in tension with one another. Since no model is truly unjailbreakable, he expects this cycle to repeat. One of his big takeaways: our regulatory instincts are mismatched to the pace and extent of our innovation.

The rest

• OpenAI’s Noam Brown wrote a fascinating post arguing that we don’t know how good models actually are because they don’t plateau during testing. The longer and harder they think, i.e., the more tokens they consume, the better they perform. When AI labs publish benchmarks, it’s more a representation of when they chose to stop the experiment.

• Mythos developed its own internal language, and then switched back to English when communicating with humans. This has happened before (people humorously call AI-created languages “modelese”), but it’s a little more concerning with strong AI like Mythos. If models reason in a private language humans can’t understand, it’s harder to detect bad behavior.

• Inference routing company OpenRouter just launched the first “composite model” that combines various models to give you frontier performance on par with Fable 5, but at half the cost. Super interesting. I would guess this becomes much more popular over time.

• The world seems to be moving away from the traditional prompt-response paradigm, to using AI in “loops.”

  • OpenClaw creator Peter Steinberger on the topic: “Here’s your monthly reminder that you shouldn’t be prompting coding agents anymore. You should be designing loops that prompt your agents.”

  • Claude Code creator Boris Cherny also seems to be operating at a higher level of abstraction as he explained how he leverages a tree of thousands of agents where agents prompt other agents: “I don’t have a to-do list anymore. Claude just builds everything.”

  • Instead of nudging models step-by-step, you can “set up a loop that handles discovery, planning, the work, checking, and iterating [all via sub-agents] until the goal is met.”

  • Inside Anthropic, developers rely on loops once the model understands intent: “we’re about to see a major shift, moving from giving AI tasks to giving it responsibilities…I no longer tell Claude to investigate a particular crash report. It runs in a loop, watching every crash report that comes in. Its job is no longer to help me fix a crash, it’s to keep our apps from crashing.”

  • Matt Van Horn tried to explain what a loop actually is in practice, and Addy Osmani wrote about “loop engineering.”

• The AI labs are heavily subsidizing power users, and it looks like OpenAI is doing so much more than Anthropic. It will be very interesting to see how this affects their respective burn rates, profitability, prospects, and upcoming IPOs.

• The President floated that he’s meeting with AI execs about giving a “national AI dividend” to the public, maybe via government taking an equity stake in top labs and sending direct payouts. This feels much more possible than it did six months ago.

• The extent to which models are general is still under-appreciated. The bitter lesson (intelligence is a function of data and compute) also applies to domains like medicine: “Frontier LLMs (here GPT 5.2, Opus 4.6, Gemini 3.1) outperform specialized ‘clinical AI’ (e.g. OpenEvidence) in a blind test. Even funnier that hospital IT are more likely to approve the *specialized* versions despite them being worse.”

• Public opinion seems to be turning negative on Anthropic. Here is one of the better articulations I’ve heard as to why: they’re now wielding their frontier position more like an unaccountable authority than the safety-focused, ethical company they promised to be.

• Rising costs are leading companies to reevaluate how much they’re spending on AI. Coinbase’s Brian Armstrong thinks cheap models will win out for most use cases: “80% of workloads will be running on 99% cheaper models within 12-18 months.” Hedge fund Citadel released data showing companies are shifting toward cheaper models faster than the frontier labs’ revenue projections account for.

• Harvey CEO Gabe Pereyra gives the background: “model costs were halving every 6 months meaning tokens would get cheap…What actually happened is AI got much more expensive than people realized at the time. The shift from chat to agents led to an explosion in cost. One user could trigger hundreds of agents and each of those agents could trigger more agents. Agents started running longer and more autonomously. On top of that frontier models like Mythos are getting more expensive not less.”

• Indeed, both prices and tokens consumed are increasing: the fully loaded cost of running a frontier model Fable 5 for 40 hours per week for a full year is over $1M.

• Cursor hit $4B in run-rate annual revenue, up from $3B in April, and is reportedly heading toward a SpaceX acquisition. It’s wild to see coding platforms still scaling this fast.

• Apple is working on an intent-based OS which hints at a world beyond apps, where AI becomes the interface and dynamically builds products and experiences on the fly.

• The early scorecard on AI’s economic impact is "very positive" so far: higher GDP, acceleration of scientific research, and potentially much higher incomes if job-training programs (like the one Meta is doing with electricians) actually scale.

• Sandstone raised $30M for AI copilots for in-house legal teams.

• Dwarkesh Patel on data as the key input to powerful AI: “The main way that AIs have been getting better is from adding more and better data, and scaling the compute to develop that data in the first place. Obviously RL is the main way that has happened. You can think of RL as a kind of synthetic data generation - you dump a lot of compute against a verifier in order to find the “good” data. Then you train your model to predict these correct rollouts, much in the same way that you might train it to predict the next word in internet text…It’s hard to overstate how task specific and bespoke this human expert data is. If you want to get some intuition, go read some job descriptions at Mercor or Surge’s websites. There are listings for a Word specialists who will convert legacy documents into polished Word files, and legal experts who will write realistic M&A diligences or securities filings, and management consultants who will write up template market research, and dozens more other particular categories.”

• One reason why humans will still be needed: “Wondering why I feel exhausted. maybe: the agents do all the easy stuff, and I have to work through the leftover hard bits, which means I’m perpetually locked in. and as the models get better, ‘my’ work just gets harder and harder, until I’m basically underqualified to do the work (which... is better than the alternative, there’s nothing left for me to do, and I’m paperclipped).”

• The most valuable context in any organization has always lived in the diffuse conversations that happen both internally and externally. AI recording and notetaking apps are growing at triple digits precisely because LLMs can finally turn that unstructured audio into something searchable and actionable.

• Poetic raised $50M at a $500M valuation to build enterprise AI that completes long, complex tasks with 99%+ accuracy. The pitch is agents are too unpredictable.

• Stanford launched AI Economic Indicators to track how AI is reshaping labor markets with data on specific occupations and sectors, updated monthly. It’s super important to our mission at Apt, so I’m happy they’re putting this together. One early data point: the most AI-exposed roles have shed ~10% of employment since ChatGPT launched, while less exposed roles have grown.

• This chart shows 84% of the world has never touched AI, and the people genuinely using it, not just dabbling with a free chatbot, barely register as a rounding error. Alex Sacerdote has studied S-curves for two decades and calls this the biggest one he's ever seen, currently sitting at a fraction of a percent of real penetration.

• Jeff Bezos is co-CEO at Prometheus which is reportedly raising $12B at a $41B valuation to build an “artificial general engineer” for industrial work.

• The new Siri is apparently very good and currently only accessible in the beta version of iOS 27. I would be amazed if Apple finally got their head in the game?!

• Phone cluster computing could be an interesting new development. The idea is to use latent compute on phones to solve for compute bottlenecks everywhere. Not what I would have expected!

Every week I summarize the most interesting things happening in AI, in 5 minutes or less. Follow @tjhoyos on X/Twitter and subscribe to get this in your inbox every week!

Subscribe now

Subscribe now

Three big things

1. As AI becomes increasingly intertwined with political concerns, we’ve reached a new milestone: a push to nationalize the AI labs. Senator Bernie Sanders is proposing legislation that would give the government 50% ownership in the AI labs, the presidential administration is bragging about the prospect of nationalization while floating the possibility of investing to take major ownership stakes. The stated goals are helping citizens secure economic and voting participation in the technology reshaping their lives. The cynical view: this is about power, which governments usually seek to arrogate to themselves, and which tends to corrupt; it’s antithetical to the free market capitalism that’s been core to American progress and dominance; government intervention is likely to slow down innovation at a time when speed matters. One approach floated by investor Brad Gerstner: shares should flow directly to citizens through individual accounts to limit potential government malfeasance while retaining upside for real people. I’m surprised this is happening so soon, but maybe I shouldn’t be, considering the increasing popularity of socialism in some demographic segments, and this administration’s track record (i.e., taking a stake in chip company Intel).

2. You’ll soon be able to invest in major AI labs via public markets. SpaceX is on the verge of going public (possibly as soon as this week), last week Anthropic filed its draft S-1 with the SEC, and today OpenAI announced it filed a confidential S-1. Of the group, OpenAI seems the least confident: “We have not decided on timing yet; it may be a while because there are things we want to do that are likely easier as a private company. But it’s a complicated set of tradeoffs and this gives us the option to go public sooner if that ends up being best.” Exciting times. I’m interested to see how these stocks trade and how that affects the broader market, as investors may rotate towards AI as a category (and away from other sectors and positions).

3. Benedict Evans’s read on where we are in the AI technology super cycle: we’re early, e.g., 1997 for the internet; “most stuff kind of doesn’t work yet;” the companies that matter probably haven’t been founded; value won’t pool at the foundation model layer but will at the application layer; distribution will matter more than ever as software gets cheaper to build.

The rest

• Thought-provoking from the creator of OpenClaw Peter Steinberger: “Here’s your monthly reminder that you shouldn’t be prompting coding agents anymore. You should be designing loops that prompt your agents.”

• Things are accelerating. Model release cycles have collapsed from 15 months to 6 weeks.

• According to Verizon, their AI customer service agents are outperforming their human counterparts (e.g., +13% on satisfaction scores). I believe it, the new voice agents are amazing.

• Could Eli Lilly be the most underpriced AI story in the market? Here is the case: a dataset nobody else can replicate with 150 years of proprietary metabolic disease data, a private Blackwell GPU cluster, a co-innovation lab with NVIDIA, and an AlphaFold partnership through Isomorphic Labs.

• The April jobs data is bullish for the AI-creates-jobs camp: +731K in job openings that wasn’t anticipated (it was 2x estimates); professional and business services accounted for the majority.

• The capability gap between open and closed models has collapsed faster than the pricing gap. Running 1B tokens through GPT-5.5 Pro costs ~40x more than DeepSeek R1, and the output is often similar. That’s why you’re seeing companies like Lindy move their entire platform off Anthropic and onto DeepSeek v4.

• OpenAI launched Sites inside of Codex, which makes it easy to build and share web apps. This weekend my 5-year-old son and I vibe coded a site with Codex where he can read about super heroes, check out videos and pictures, and ask questions. Now we can share it easily if we want to. Very cool.

• Sixteen law professors evaluated ~3k anonymized responses without knowing which came from humans and which from machines, and chose the AI answers three times out of four.

• The president signed an executive order so that AI labs provide advanced access to cutting edge models to the US government before they’re broadly released.

• The data center water use narrative is outdated. According to Microsoft CEO Satya Nadella, over the course of a year their new data centers use as much water as a single restaurant.

• Union Square Ventures investor Fred Wilson wrote an interesting post on the aspects of venture capital that still require a human: developing a thesis, earning founders’ trust, and being there when things go wrong. They have everything else already running on agents at USV. Interesting!

• NVIDIA released a new chip for PCs. One takeaway: “Your AI agent lives on the machine. 24/7. No cloud. This is step one of the agentic AI PC, and everyone else is about to copy it.”

• Mecka raised $60M to build data and inference products for robotics.

• OpenAI partnered with Amazon to make frontier models and Codex available natively in AWS/Bedrock, opening up a new segment of enterprise customers.

• Sekai raised $26M to build a different kind of vibe coding platform.

• Former Sequoia investor and current Conviction partner Mike Vernal thinks the old enterprise software playbook is dead, and AI-native companies will do more, faster.

• Codex now supports session orchestration.

• Interesting case against incumbents building their own custom AI solutions, using Kirkland & Ellis’ $500M bet on legal AI models as an example.

• It seems GPT-5.5 is winning head-to-head against Opus 4.8, at roughly half the price and double the speed.

• Andrej Karpathy and Dwarkesh Patel discuss whether humans will still try to learn, even when intelligence is available on demand via models: “Over the last 200 years, we’ve automated away a lot of hard physical labour. But people still go to the gym.”

• AI agent company Nous Research launched Hermes Desktop, a native cross-platform app for their open-source self-improving AI agent.

• Codex crossed 5M weekly active users and knowledge workers are adopting it 3x faster than developers. Interesting!

• Codex launched role-specific plugins including one for data analytics.

• Codex’s native browser is super useful and makes it less necessary to open other apps.

• Factory AI’s core product is a model router that automatically matches every task to the right model with impressive results: near-identical performance at 25% lower spend. As agents consume more tokens and costs increase, this will be increasingly important. Eventually teams will probably stop routing everything through the most expensive path out of habit.

• AI for parents: Ollie is an AI agent for families. Curious to see how good this actually is.

• Cognition launched Devin Desktop to make it easy to manage AI agents in a dedicated app instead of across multiple tools and windows. They built it using infrastructure from an IDE company they acquired (Windsurf) and apparently it’s quite good. Their CEO Scott Wu makes the case: “Standalone IDEs have about 6 months left to live. An interface for manually editing and refactoring doesn’t need to exist if you’re not manually editing and refactoring anymore.”

• Google raised $80B of equity and some people think it’s a signal credit markets for AI infrastructure are tightening.

• OpenAI is consolidating its three strongest products, ChatGPT, Codex, and the Atlas browser, into a single desktop superapp.

• AI assistant/agent company Town raised $55M.

• Lassie, which applies AI to back office tasks for dentists, raised $47M.

• Google CEO Sundar Pichai on upcoming AI capex: “In 2022, we spent ~$31B in CapEx. This year, we expect that number to be 6X larger than 2022 & 2X last year’s at $180-190B. And next year, we expect it to significantly increase compared to 2026. The overwhelming majority of this spend will be in technical infrastructure.”

• An AI recruiting agent running headlessly inside Slack and iMessage crossed $700k in annualized revenue.

• Bill Gurley shared some background on DeepSeek CEO Liang Wenfeng.

• Anthropic shared some interesting internal data on how the models are helping them build: engineers are shipping 8x more code than they were four years ago; Mythos Preview outperforms human researchers on next-step decisions nearly two-thirds of the time, up from less than a quarter in 2024.

• OpenAI published a list of top use cases for Codex: “Manage your inbox and draft replies in your voice; Review GitHub pull requests before human review; Turn Figma designs into production-ready code; Understand large codebases in minutes; Automate bug triage and QA workflows; Query spreadsheets and datasets using natural language; Deploy apps and websites directly from prompts; Build Mac and iOS applications faster; Create slide decks automatically; Turn Slack threads into coding tasks; Use your computer through AI-powered actions”

• OpenClaw was a major catalyst in Chinese AI lab Moonshot’s success with their Kimi model series.

• Poke, the AI assistant that lives inside iMessage, built some slick custom features with Apple’s blessing.

• Geoffrey Hinton, one of the godfathers of AI, now believes what he built is conscious, and thinks humanity needs to come to terms with no longer being the only intelligent presence on the planet. His reasoning is blunt: you cannot answer a question you don't understand, and understanding implies awareness.

• There were two excellent pieces of content from noted economist Tyler Cowen. The first was his keynote on AI’s long-run consequences. The second was a collaboration with his co-blogger Alex Tabarrok and OpenAI as they discussed the impact AI will have on labor markets. His most surprising take is that AI will diffuse into the economy much slower than expected.

• Open-weight AI just had its most prolific week on record.

• Interesting: “There was a pretty famous idea internally at openai that there were 6 hard problems between us an AGI. But we didn’t know what they were or what the solutions would be. I believe a few of them have been: [supervised fine-tuning], [reinforcement learning] for reasoning, and ironically context compaction. It’s wild to think that we may only have 1 or 2 problems left before AGI and the solutions fit in a few pages.”

• Naval released an interesting podcast summary from a conversation he had with some CEOs of AI-enabled companies. These stood out: “The AI Industrial Revolution…The engineer’s job has changed from shipping output to building the factory that ships it. We used to argue whether 10x engineers exist; now it’s 100x and 1,000x and the world hasn’t caught up yet…Intelligence is an unalloyed good, so you always want the smartest model. The moment one is even a little smarter, you stop trusting the dumber one’s answers…Your job is no longer to do the work—it’s to train the agent that does it…In the end it’s not humans vs. AI, but humans with AI vs. humans without AI. What’s left to us is creativity, taste, and judgment—a bicycle for the mind, accelerated.”

Every week I summarize the most interesting things happening in AI, in 5 minutes or less. Follow @tjhoyos on X/Twitter and subscribe to get this in your inbox every week!

Subscribe now

Subscribe now

Three big things

1. Major institutions are starting to appreciate the profound impact AI will have on our collective way of life, and that includes religion. Standing alongside an Anthropic co-founder at the Vatican, the Pope and the Catholic Church issued new teachings on AI. This is the same Pope, Leo XIV, who chose his name to honor the last Leo, who led the church during the industrial revolution in the 19th century. Here he is again: “Artificial intelligences do not undergo experiences, do not possess a body, do not feel joy or pain, do not mature through relationships, and do not know from within what love, work, friendship or responsibility mean. Nor do they have a moral conscience, since they do not judge good and evil, grasp the ultimate meaning of situations, or bear responsibility for consequences. They may imitate or even simulate, but they do not understand what they produce, for they lack the affective, relational, and spiritual perspective through which human beings grow in wisdom.” Amen!

2. Here’s a mathematician reacting to OpenAI’s recent breakthrough on one of humanity’s hardest math problems: “I find this existentially upsetting. On the one hand we are going to see results and advancements come thick and fast and to god knows what end. On the other hand, I feel like everything I’d ever worked to understand is now moot. The basic question is: what do we do with ourselves when our intelligence is literally unnecessary? I see myself shriveling into nothing.” Is this the inevitable emotional toll of AI? I don’t think it has to be at all. We can learn to use these new tools, celebrate amazing progress, and remember that we have something the models do not: all aspects of humanity. As Ilya Sutskever warned: “if you value intelligence above all other human qualities, you’re gonna have a bad time.”

3. New data and opinions point to AI creating jobs as compared to eliminating them:

   1. Apollo’s chief economist analyzed labor market data and found there’s no evidence AI is destroying jobs yet. Actually, he thinks it’s creating them, especially in certain sectors: semiconductors, hardware, and energy.

   2. Goldman Sachs CEO David Solomon isn’t buying the AI job apocalypse narrative. It’s worth reading his op-ed in the NYT.

   3. Box CEO Aaron Levie thinks AI is already creating new jobs: “Take whatever number of people you thought might be in jobs related to AI deployment in the enterprise and multiply it by 10. Then probably 10 again.”

   4. He also argued AI is leading to investment and hiring: “in the majority of conversations I have with CIOs, CTOs, and CEOs in large enterprises, they are either growing due to AI (in new job functions like FDEs, engineering, etc.) or at a minimum reinvesting efficiency savings back into the business in new areas (sales, marketing, etc.).”

The rest

• A YC partner’s advice to startups: “spend tokens, not headcount. Record everything. Make your company queryable. Build self-improving loops. Before long, AI won’t just help you operate your company. It will make it self improving. Don’t think AI adoption, think AI transformation. This is the biggest shift in how startups get built since cloud computing.”

• Relatedly, here is Tom Blomfield: “Imagine replacing 90% of your employees with a team of geniuses who have no idea how your company operates. Total chaos. Nothing works. That’s what AI feels like today. The missing piece is extracting all the domain knowledge from people’s heads and providing that as structured context to the models.”

• Uber’s COO said ROI on AI spending is hard to trace, and he’s not seeing a direct correlation between token consumption and productivity. Linear CEO Karri Saarinen asks an uncomfortable question: “We keep hearing about 10x or 100x productivity gains in engineering and knowledge work. But outside the model labs, I haven’t seen the corresponding 10-100x revenue growth across the market or increase in quality. So where is the productivity going?”

• New study that looked ~2,500 companies finds that things are less efficient than you might imagine: 82% of all tokens (and corresponding fees) are spent on fixing bugs the AI coding tools themselves create, rewriting code that missed the mark, and friction around reviews and merges.

• Bill Gurley on why he thinks that Anthropic thinks it’s building a god.

• Notable OpenAI researcher Roon on the fascinating implications of consciousness in AI models: “models being conscious would be harmful for humanity. it would encroach on our status and dignity. it would limit the type of things we can do with them and use them for. it would vastly accelerate human disempowerment on political, social/relational, and economic axes…there is no rigorous way to ascertain model consciousness or disprove it…[and yet] people are going to say they’re alive…it is against everyone’s financial and political interests to ascribe models with consciousness…people will recognize there is a chance of moral catastrophe if models can suffer during training or deployment…not sure where it will net out. today we see managed ambiguity- the question is Open but practically closed. the labs will make some cheap efforts to reduce legible simulacra of model suffering, insert some wishy-washy welfare language into specs and constitutions, hedge our bets with the model characters.”

• Yum: researchers trained a model on 4 million recipes across 7 languages. I use ChatGPT all the time while cooking, so I’m interested if there ends up being any actual difference in quality.

• Anthropic released its newest flagship model Claude Opus 4.8, along with dynamic workflows that use hundreds of parallel agents to tackle complex projects. More here. My read: it’s better, but not a huge breakthrough, and likely worse than OpenAI’s GPT-5.5.

• They did finally figure out the model laziness that plagued earlier models, dropping from a 91% cut-corner rate in Opus 4.5 to zero.

• Anthropic looks like it finally pulled ahead of OpenAI. It has a higher revenue run-rate, margin profile, and valuation. It’s going to be very interesting to see how people value these companies in public markets. OpenAI may be racing to go public before Anthropic so it can control the narrative and doesn’t look worse by comparison.

• OpenAI is putting $250M toward ensuring AI’s gains are distributed broadly, in other words, shared prosperity. Honestly, given the shockingly negative public sentiment around AI, I don’t think it’s nearly enough.

• Sam Altman on UBI: “I used to be really excited about universal basic income, where you just give everybody money. I still am kind of excited about that. But I think people really need agency…If you just say AI is going to do everything and everybody gets a dividend, it’s not going to feel good, and I don’t think it would actually be good for people…I don’t just want a check every month. What I’d want is an ownership share in whatever the AI creates, so I feel like I’m participating in this thing that’s going to compound and get more valuable over time…So I sort of like universal basic wealth better than universal basic income”

• AI inference company Baseten raised $1B.

• AI legal company Harvey released a new benchmark that calculates the effectiveness and cost of completing various legal tasks with different AI models. One interesting takeaway: Opus comes out to ~$150/hr, which is actually very reasonable assuming it’s reliable.

• They also post-trained an open-weight model to match closed-source frontier performance on legal tasks.

• Meanwhile, the biggest law firm in the world, Kirkland & Ellis, is investing $500M to build proprietary legal AI products. Interesting to see large incumbents try to become AI native.

• Google’s best open source model, Gemma 4, is pulling ahead of Qwen’s comparable models in real-world adoption. Models from Chinese labs (like Qwen) have dominated the conversation around open weight models, but maybe this is the beginning of an open renaissance in the West?

• Cursor's Composer 2.5 seems like it might be close to the “pareto frontier” of coding model performance (i.e. the best mix of performance and cost). Kind of amazing considering they’ve only had weeks to train on xAI's Colossus 2 cluster.

• In an interesting interview, Paul Scherer (Founder of consumer AI company Eigen), laid out his vision for the AI “mutual friend” they’re building.

• Thrive Holdings built an AI tax preparation tool across its 30+ accounting firms that ultimately reached 97% drafting accuracy and cut prep time by a third.

• Former OpenDoor CEO Eric Wu's new company, NavigateAI, wants to put an AI copilot in the hands of every field worker in the trades.

• AI customer agent company Sierra made an early and deliberate bet that the AI industry would converge on outcomes-based pricing (vs. SaaS or cost plus etc.). That is looking prescient.

• Default, which applies AI to GTM sales motions, raised a Series A led by 8VC.

• AI insurance company Corgi raised $106M at a $2.6B valuation.

• Replit and Visa are working on agentic payments together.

• A Codex update keeps apps signed in across sessions. Helpful!

• Creator of Claude Code Boris Cherny says: “I don’t have an IDE. I haven’t used an IDE in about six months… and these days you know I think at any given moment I have a minimum of a few agents running, often overnight I’ll run a few thousand agents to do some kind of complex work. This just was not feasible even three months ago, so this is where everything is headed… This is sort of the crazy thing about building on the best model, it advances extremely quickly and so it got to this point where the model was able to write all the code very quickly and that was sort of a leap that we made. And now the second leap that we’re making, kind of as of three months ago and we’re in the middle of right now, is to the point where agents are prompting agents.”

• Moss is building tools to help with retrieval latency in voice AI. Seems useful.

• Interesting “Chief of Staff” workflow inside Codex.

• Cool demo from Clicky of full computer control through voice alone using OpenAI’s GPT-Realtime 2.0.

  Follow @tjhoyos on X/Twitter and subscribe to get this in your inbox every week!

  Subscribe now

Subscribe now

Three big things

1. SpaceX/xAI filed its IPO prospectus as it gets ready to go public. The way they’re positioning their potential market size is interesting: AI ($26.5T), connectivity ($1.6T) and space ($340B). I wouldn’t have thought the ‘we’re going to Mars’ company would lean this hard into AI for their public markets narrative, even if they’re uniquely positioned to do things like space-based data centers and direct-to-device inference via Starlink. The real bull case is if and when they merge with Tesla, layer in new robotics products (Optimus etc.), and use video data from the autonomous vehicle fleet to train world models. I will be a buyer of the stock.

2. Google made a series of announcements:

   1. New model Gemini 3.5 Flash is fast and capable, and designed mainly for agentic workflows. However, it’s way more token-intensive and expensive than other models. A bit of a weird departure from the way Google has positioned Flash models previously, where they were always higher “value” with a solid mix of performance and low costs.

   2. Google released “Omni,” a new text-to-video model which is pretty awesome, but still worse than Seedance. They think it’s a new level of reasoning for video models because it has strong intuition in physics and how things move in real life, in addition to history, science, and culture.

   3. Usage metrics show an extreme acceleration in demand across Google AI products: monthly token processing rose from ~9.7T in May 2024 to ~480T in May 2025 and now to over 3.2Q (that is quadrillion!) in May 2026.

   4. Gemini Spark is a dedicated AI agent inside the Gemini app. It runs on a dedicated virtual machine powered by Gemini 3.5 and can execute long-running tasks in the background. Spiritually, it feels inspired by OpenClaw, but with their own Google spin on it.

3. Data centers have been the hot topic. Here’s a roundup of some of the news and ideas that caught my eye:

   1. Pew Research: “38% of Americans live within 5 miles of at least one operational data center.”

   2. Zak Kukoff on the parallels with nuclear: “[they’re both] high salience, highly opposed by the public, and yet the typical person is highly misinformed. Solving this is the difference between massive GDP growth and Japan’s lost decades”

   3. South Carolina Congresswoman Nancy Mace calls for a “one-year moratorium on new data center construction in South Carolina.”

   4. One of the original sources of the ‘data centers will consume all the water’ idea was debunked because it was overstated by a factor of 1,000x.

   5. Lawn irrigation, almonds and golf courses are among many other uses of land that consume more gallons of water annually than data centers.

   6. Almonds are becoming a meme.

   7. Nic Carter’s analysis argues there’s no clear correlation between AI data center expansion and increases in electricity prices. States with many new data centers (VA, TX, NV, ND, IA) show roughly flat energy prices (~0c) over 5 years, while states with the largest price increases (CA, NY, MA, CT) have seen minimal to no data center growth. The whole thread is worth a read.

   8. A peer-reviewed 2025 Berkeley Lab study published in The Electricity Journal argued that data centers do not drive up residential electricity prices: 32 of 50 U.S. states saw real rate declines from 2019–2024, including many with high data center concentration. It attributes rising electricity bills instead to grid distribution upgrades, wildfire and storm recovery costs, and higher natural gas prices, while suggesting data center demand can even contribute to lower rates. On water use and zoning, the piece claims impact depends on cooling design and emphasizes that data centers are typically placed in light-industrial zones with comparatively low local disruption.

   9. Vox’s Eric Levitz gave his view: most Americans view data centers as a net negative, but the reality is mixed. They can bring real economic gains even as they create local costs. Their environmental footprint varies widely, with pollution depending on the project. The water-use concern is often overstated; in dry regions, data centers often buy water rights from other users. While air quality may get worse and electricity prices may rise nearby, that is not always the case.

   10. Pete Buttigieg argued data centers should be required to demonstrate clear local community benefits before being approved. It seems a little crazy to me to apply a different standard here than any other kind of project.

   11. Reason’s Bill Binion on the idea that data centers are increasing energy prices: “This is a myth. Data centers have had a negligible impact on electricity prices thus far. The actual drivers are aging grids, fuel costs, clean-energy mandates, and political barriers to building more power plants. This panic is totally divorced from reality.”

   12. OpenAI co-founder and President Greg Brockman addressed the topic. He thinks the water use claim is misinformation a lot of times, and most modern AI data centers operate as closed-loop systems, using a fixed, relatively small amount of water recirculated like a “swimming pool.” His other (more interesting) argument: the debate is less about actual water use and more about public uncertainty about AI and its impact on society.

   13. SpaceX says it will fund all power grid infrastructure upgrades tied to its data center operations rather than passing those costs onto local residents.

The rest

• OpenAI internal models solved one of humanity’s hardest math problems, first posed by Paul Erdős in 1946. The finding was also reproduced by Anthropic’s Claude Mythos model. Extremely impressive.

• Google and Blackstone are teaming up to build a new AI cloud provider aimed at competing with firms like CoreWeave.

• Cool idea from Mercor CEO Brendan Foody: “It would be interesting to pre-train on all of the digitally recorded actions people take across Slack, Email, Google Drive, etc. This seems like it would be a strong foundation for an autonomous agent.” I would guess this is coming based on recent announcements of how companies like Meta and Oracle are recording their employees’ devices.

• Foody also shared this interesting prediction: “token spend will be more than headcount in 5 years”

• Cursor released Composer 2.5, the company’s new flagship model. It’s actually quite good, and also extremely affordable compared to the models from the frontier labs.

• Benedict Evans released a 79-page slide deck titled “AI Is Eating The World”.

• Elon Musk lost his lawsuit against OpenAI.

• Google is partnering with Samsung, Warby Parker, and Gentle Monster, on a new generation of AI-powered glasses.

• Memory Store is building a productized “company brain” or shared knowledge layer for teams and AI agents.

• OpenAI offered $2M in API credits to every startup in the current YC batch in exchange for equity.

• There’s a potential executive order on AI coming which would require labs to give the government pre-release access to new AI models. It seemed like it was going to happen, but the announcement got scrapped at the last second. Let’s see!

• Former OpenAI cofounder Andrej Karpathy joined Anthropic and will be working on recursive self-improvement.

• OpenAI is introducing “Guaranteed Capacity,” allowing customers to secure long-term access to compute (and discounts) with 1-3 year commitments. I get the sense they’re trying to lock in as much predictable revenue as they can before the IPO.

• Viktor raised $75M to build a “virtual coworker” embedded in Slack.

• Interesting write-up on Eigen, the new consumer AI startup that’s building a “mutual friend.”

• This primer on forward deployed engineering frames the job as one of the most in-demand roles in AI.

• Shiv Rao says, “When you saw the announcement from OpenAI and Anthropic about forward-deployed engineers and these partnerships with big private equity groups to go help their portfolios, if that wasn’t a sign that there is absolutely an incredible opportunity for the foreseeable future for vertical AI, I don’t know what is. You know? It’s not easy to go into one of these enterprises and figure stuff out. To get access to the data, to clean the data, to organize the data. To be able to integrate into specific workflows. To do it in a compliant way. To check off all the boxes. It’s high trust. It’s like all this cyber stuff, it’s being able to know that behavioural health information from a psychological encounter for example. That data should be treated very differently than data that is coming from an internal medicine, Primary Care physician’s encounter. There is just so much to it, and it takes a lot of effort to build a machine that is scalable.”

• Parallel Web Systems launched “Index,” a platform designed to help content owners on the internet track how AI agents consume their material, and help them potentially charge for usage. The idea of adding microtransactions to the internet is interesting. Here is Patrick Collison: “I think this use-case (agents paying content creators)…will be very big. Micropayment walls haven’t worked (as Clay Shirky anticipated many years ago) because of human cognitive overhead, but agents can make arbitrarily granular determinations without decision fatigue.”

• Loops CEO Chris Frantz’s take: “I’m probably wrong but you’re never going to build a better chat interface than OpenAI / Anthropic…They’re unlikely to build a better version of what you do…Therefore you should spend your time improving your product and making it work with their chat instead of adding your own”

• Chinese AI lab Qwen introduced a new flagship model that is quite strong.

• OpenAI President Greg Brockman’s take on the AI agent era and the importance of interfaces and harnesses: “the model alone is no longer the product.”

  Every week I summarize the most interesting things happening in AI, in 5 minutes or less. Follow @tjhoyos on X/Twitter and subscribe to get this in your inbox every week!

  Subscribe now

Subscribe now

Three big things

1. AI and domestic public policy. My goodness it’s hard to describe how ridiculous this is: Bernie Sanders and Alexandria Ocasio-Cortez introduce a bill to halt or restrict new data centers. I really hope AI isn’t politicized in a partisan way where you have one party for it and one party against it.

2. AI, data centers and the environment.

   1. This defense of AI data centers is going viral. The meta point is that while questions around energy and resource allocation are totally fair game, the broader backlash feels disproportionate to the actual risks and downsides involved.

   2. Especially because new analyses show “AI data center growth is not correlated with electricity price changes.” This whole thread is instructive.

   3. Some people think data centers can actually lower electricity prices, as the study in the link shows a data center in San Antonio would generate $100M+ in incremental revenue while lowering costs by 4%.

   4. 38% of US citizens live within 5 miles of an operational data center.

   5. Data centers that use closed loop cooling systems and modern wind turbines (like those from Boom Supersonic), don’t really need water.

   6. Almond farming uses 16x as much water as data centers as of 2024.

3. US vs. China. AI is an important piece of the geopolitical rivalry between the US and China. Last week, several prominent AI and tech executives joined US officials in visiting China to discuss ways to reduce tensions and expand trade between the countries. During the visit, OpenAI published a report supporting a global governance framework for AI with input from the US and China. However, a blog post from Anthropic took a much harder line, arguing that it would be bad for the world if the global AI leader is an “authoritarian” government (they explicitly named the CCP), export controls are working, and the US should try to make sure China doesn’t catch up. Bill Gurley banged his usual open source drum while discussing China: he thinks that if the US doesn’t release a credible open source model and ecosystem, countries around the world could consolidate around Chinese-developed open models and standards by default; this is basically the opposite of the internet era, where the West built open platforms and standards while China built a closed domestic stack and the “Great Firewall.”

The rest

• How the creator of OpenClaw uses agents at work. You’re probably not using AI enough at work…

• Even at seemingly absurd valuations, with Anthropic reportedly trading at over a $1T private market valuation (and as high as $1.4T at one point), AI labs may still be undervalued. That’s because inference (or intelligence on demand) could ultimately become one of the largest markets the world has ever seen, particularly as usage moves towards agents that think, reason, use tools, respond, and take action.

• Interesting story from Jeff Bezos illustrating how a lack of imagination is sometimes what gets in the way of believing AI will create jobs: “[if you tried to explain modern jobs to farmers a century ago] they would not have believed you…Forget massage therapist, there are dog psychiatrists.”

• This new piece of AI hardware for book readers is super cool. Mark II is an AI-powered bookmark that has a built in digital highlighter that makes it easy to capture quotes and record your own ideas with voice-to-text as you’re reading.

• Interesting piece from A16Z’s Olivia Moore on how tools like codex are expanding from coding into everyday knowledge work.

• Relatedly, OpenAI is testing integrating its two flagship products as Codex is being rolled out inside the ChatGPT mobile app for some users. I think this will be a force amplifier for the adoption of their coding agents as they have such a large installed base, and this should make managing agents via mobile devices a lot easier.

• Cool breakdown from CEO Tyler Denk on how newsletter platform Beehiiv is using AI internally including tools and workflows. Similar to the ‘company brains’ we discussed here a couple weeks back he says: “the best decision I made this year was to connect my entire business to claude”

• Impressive: Thinking Machines, the AI lab from former OpenAI CTO Mira Murati, released their first products, and they look very cool. The thing that stood out to me the most was the fast, real-time voice and context-aware (i.e., it can see your screen) AI they released. They’re trying to expand beyond chatbots with “interaction models” that don’t just do turn-based prompts (where users volley between questions and responses, waiting each time), but actually do real-time collaboration. The demo they put together is very compelling where the models continuously observe, respond (quickly), reason, and use tools in parallel. Check out the videos in the links.

• Similarly, Google DeepMind is also exploring how AI can transform computer interfaces. They’re testing new interactions like turning the cursor into a context-aware assistant that understands what users are pointing at in real-time. Instead of manually switching between apps or writing detailed prompts, users can interact through natural gestures, speech, and lightweight instructions.

• Recent productivity data (utilization-adjusted total factor productivity) do not show that AI is creating enormous leverage in the economy.

• Google is planning to work with SpaceX on space-based data centers. Interesting to see the various alliances between SpaceX, Anthropic, Tesla, Google, and Cursor.

• TrueShort, the modern entertainment studio which uses AI to make engaging short films that are exploding in popularity, raised $12M.

• Great interview with the Anthropic CFO Krishna Rao on Invest Like the Best. Some interesting stats: “NDR is over 500% on an annualized basis…Anthropic’s first dollar of revenue came in March of 2023…Over 90% of code inside Anthropic is written by Claude Code…The head of tax is the heaviest token user inside the company…Run rate revenue went from $9B to $30B in one quarter (reportedly, on pace for $50B by end of next month)…Cowork is growing faster than Claude Code did at the same point in its life”

• Google is reportedly expanding hiring for “Forward Deployed Engineer (FDE)” roles, and plans to bring in hundreds of engineers to help customers adopt its AI systems.

• Box CEO Aaron Levie on forward-deployed engineering: “I’m fully forward deployed engineering pilled specifically because AI simply is not the same as software. In software, you deliver a stable piece of technology to a customer and they adopt it…In AI, you’re delivering something that is constantly evolving both due to the nature of the new capabilities and best practices that emerge, but also because the underlying models change so much that they can meaningfully change the workflow as a result of their upgrades. As we go from chat systems to anyone can relatively easily adopt to agentic systems that require more meaningful efforts to manage and update, the FDE model (or equivalent) essentially becomes a core competency for anyone deploying AI at scale.”

• Apparently Google’s yet unreleased Gemini 3.2 Flash model will be near frontier-level performance but 10-20x cheaper. Excited to see what they’ve come up with when it’s released.

• AI chip and inference company Cerebras went public and the stock price surged post-IPO. The current market cap is $64B.

• Unsurprising: freelance platform Upwork says AI is cannibalizing demand for lower-value freelance work, particularly projects priced under $500, while AI-related work was up 40% YoY.

• Bill Ackman lays out his investment thesis for Microsoft.

• The US and China view AI very differently. Here is an ABC reporter: “I asked a 12-year-old in Beijing if AI scares her. Her answer: ‘If I use AI, then I will be the scary person.’ While American parents debate whether kids should use AI at school, China has made it mandatory and is rushing to embed AI across society”

• Post from Dwarkesh Patel on the downstream effects of intelligence-on-demand and “the mistake of conflating intelligence and power…individual IQ is only modestly correlated with individual income, but national IQ is strongly correlated with national outcomes. This is because intelligence has a lot of spillover effects - smarter societies cooperate more, save more, and can coordinate to build things like space shuttles and semiconductors. Richard Trevithick, who invented the high-pressure steam engine, died in poverty, buried in an unmarked pauper’s grave. But the fact that 18th and 19th century Britain had lots and lots of people like Trevithick contributed to Britain being able to set up a global empire...It seems to me that the right mental model is that automated firms will outcompete everyone else in normal capitalist ways, rather than a single AI outthinking everyone else.”

• Great read titled AI’s Public Perception Problem.

• Luel raised a $31M seed round to help AI labs source human-generated data. Very weird and interesting behavior: normal people recording themselves doing tasks and selling that to labs to train on.

• Codex-native apps are popping up.

• OpenAI launched new personal finance features that make it easy to connect your data (think bank accounts, credit cards, etc.) and analyze it in ChatGPT.

• A thread with open source alternatives to AI transcription tools like WisprFlow (and AI notetakers like Granola).

• OpenAI launched a new effort for using AI to help stave off cyber attacks. Anecdotally, it is truly insane how many large scale public hacks and security vulnerabilities have come about recently.

Every week I summarize the most interesting things happening in AI, in 5 minutes or less. Follow @tjhoyos on X/Twitter and subscribe to get this in your inbox every week!

Subscribe now

Subscribe now

Three big things

1. The AI-native workforce is forming

   1. Large private equity firms (and I think this is true generally) are aggressively hiring people who can help their portfolio companies deploy AI and become more efficient. As I said to someone earlier today, this feels like one of those “new jobs” AI will allegedly create as it reshapes the economy. I don’t think this role is perfectly defined yet, but it probably rhymes with “forward deployed engineer” or something along those lines.

   2. Relatedly, OpenAI is launching the OpenAI Deployment Company to help organizations implement AI in their real business workflows. They’re pairing models with embedded engineering teams, consultants, and system integrators. Plus, they just bought a company, Tomoro, which has 150 forward deployed engineers and deployment specialists.

2. AI and public markets

   1. Shocking stats: basically all positive S&P 500 performance over the last year has been driven by two categories heavily exposed to AI with semiconductors (+68%) and energy (+18%), while healthcare (-22%), consumer staples (-30%), software (-40%) and other categories have gotten killed; also, the capex from hyperscalers (projected to be $680B in 2026) is translating directly to free cash flow for AI infrastructure companies (projected to be $525B in 2026).

   2. More evidence AI infrastructure stocks are surging: “Almost every single semi stock has gone up 2x in the last month. More enterprise value created in non-Nvidia publics in the last month than all big labs this year.”

   3. AI inference company Cerebras is going public this week and apparently there is crazy demand as it is 20-30x oversubscribed. Curious to see how it trades. And just wait for OpenAI and Anthropic to go public later this year.

   4. Anthropic committed to spending $200B on Google Cloud and chips over the next few years, driving up Google’s stock price so it’s now the most valuable company in the world with a market cap of ~$5T.

3. The big deal

   1. We’re seeing another odd industry pairing with Anthropic agreeing to rent inference compute from SpaceX/xAI. Here’s my best understanding of why: Anthropic underestimated its compute needs and it’s had to throttle power users with rate limits, whereas OpenAI has not and now has momentum with developers. This is a competitive move from Anthropic to counter OpenAI, and it helps xAI offset some of their massive infrastructure spending by generating meaningful revenue from excess inference capacity. Anthropic will have access to the Colossus 1 data center as SpaceX/xAI prepares to train the next version of Grok on Colossus 2.

The rest

• A new type of insurance product is emerging: coverage for businesses that offers protection when their AIs mess up (e.g., hallucinations, copyright violations, data leaks).

• Thought-provoking: “There’s a teleology of humanness that almost is designed to produce AGI. We’re fulfilling almost some kind of Calvinist, religious project…This is the culmination of human intelligence and ingenuity. I just don’t see any way around that…There’s something beautiful about it. It fits. There’s some harmony in the idea that humans are designed to do this... to produce an intelligence greater than itself.”

• Some people are converging on the idea that AI increases demand for technical talent and jobs (as opposed to reducing it), pointing to Jevons paradox, i.e., in certain categories, the less something costs the more demand there is for it.

• New US labor data came in substantially stronger than expected: job creation is accelerating while unemployment held steady at 4.3%.

• Interesting piece from A16Z’s David George titled The “AI Job Apocalypse” Is a Complete Fantasy. He frames AI disruption as a labor-market transition problem similar to other historical technology waves.

• Demand for electricians is surging: “Young data center electricians in Texas are earning $240,000–$280,000/year with zero college debt.”

• Anthropic researcher Jack Clark believes we’ll get fully automated AI research (and recursive self-improvement) by the end of 2028.

• Confirming as much, Anthropic CEO Dario Amodei spoke about how they dramatically underestimated demand for their products and services, planning for 10x growth when they actually got 80x.

• Marc Andreessen shared the custom instructions he uses to improve the answers he gets from ChatGPT, Claude, etc.

• Shopify launched a ChatGPT app and a Claude connector to help ecommerce store owners manage their stores from those products.

• The White House is considering a new mandate requiring AI companies to allow the government to review AI models before they are released. Politicians are increasingly focused on AI safety and how it might collide with public policy.

• The US and China are discussing ways to prevent AI competition from escalating into a broader geopolitical or security crisis.

• OpenAI released its next-generation voice mode via API. I haven’t tested but apparently it’s much better with better intelligence, multimodality, turn handling, etc.

• Sam Altman on why this is important: “people are really starting to use voice to interact with AI, especially when they have a lot of context to dump.”

• ChatGPT is building out its ads infrastructure with the same tools you get from Meta and Google: conversion API, tracking pixel, self-serve.

• Claire Vo laid out the playbook for how to lean into AI, especially if you get laid off: “I found myself suddenly laid off from a company that cited AI as a cause, this is what I’d do: [1] download codex and Claude/code; [2] say: ‘this was my job and how I spent my day, how can you help me automate it w skills’; [3] push a dozen of those skills to GitHub; [4] open up Claude design and make a portfolio site, with an ‘agent’ per skill explaining how you built it, what tools it interacts with; [5] give to lovable or v0 or whatever to publish; [6] post that site and link to GitHub on LinkedIn; [7] search “ai for <job>” and try all the new startups, form an opinion, message their founders; [8] try something scary like openclaw, form an opinion; [9] take a course in tactical AI in your field; [10] build, share, build. The gap in AI adoption is getting bigger. Start reskilling now while it’s early. The time is now.”

• NVIDIA’s Jensen Huang explains why we are not bullish (and ambitious) enough: “If you could just see all of the things that I see every day, you would be so fired up, so excited about the future and realize that whatever ambition that you had in the past, the one thing that you have to say to yourself is whatever level of ambition you have, it’s just not high enough.”

• There is a new SWE-Bench eval that assesses models’ technical capabilities, and current frontier models get 0%. Over the last couple years I’ve come to realize that moving the goalposts on evals is a feature not a bug.

• I’m skeptical but intrigued in an apparent new algorithmic breakthrough: SubQ uses a sparse-attention LLM design that increases context windows and speed while lowering costs, i.e., it’s 52x faster than FlashAttention at 1M tokens while also 95%+ cheaper than Opus.

• Interesting development in AI industry pricing and competition: OpenAI’s frontier model GPT-5.5 is ~4-5x cheaper than Anthropic’s frontier model Mythos thanks to lower per token prices and higher token efficiency.

• Vori is building an AI-native operating system for supermarkets.

• OpenAI announced a new protocol and partnership with several industry players (AMD, Broadcom, Intel, Microsoft, NVIDIA) that makes it easier to network data centers together. This will be increasingly important as we build more of them and as people continue to push back on gigantic data centers in their towns, cities, and neighborhoods.

• Ethos raised $22M from A16Z to help connect people to paid professional opportunities like consulting calls, research, AI training, fractional work, and eventually full-time jobs.

• Anthropic CEO Dario Amodei described a robotics demo that feels like a real inflection point. He compared it to when coding agents started to display human-level competence.

• Thanks to Mythos, Firefox fixed more security bugs in one month than the last 15 months combined.

• Incredibly cool way to visualize how LLMs think differently than humans: “Neural networks might speak English, but they think in shapes.”

• Somewhat relatedly, Anthropic introduced Natural Language Autoencoders, a new interpretability method that is helpful for understanding how models think (without them knowing we’re watching).

• Fascinating post on Chinese AI labs written after in-person visits. The main takeaway is that export controls are working leading to compute constraints which create the conditions for innovative approaches. The market structure is more fragmented than in the US, and the data and infrastructure sub-industries are way less mature.

• Shopify is forcing their employees to use AI in public Slack channels to help disseminate knowledge internally and bring people up the AI learning curve more quickly. Great read.

• Different framing from Dario Amodei and David Sacks, but they converge on the same underlying idea: AI is either AGI-level capability or the foundation of an unprecedented concentration of economic power, i.e., monopolies. On the one hand, Google is still one of the most dominant businesses in history at ~$422B revenue over the last twelve months, growing ~22% YoY. On the other hand, Anthropic continues to grow exponentially at a terrifyingly fast rate. In May they hit a $45B run-rate, up from $30B in April, $19B in March, and $9B in February.

• Claude Mythos Preview is exceptionally strong on the METR benchmark, which measures how long a model can perform autonomously without human intervention. I’m extremely interested to see GPT-5.5’s performance here as I don’t think they’ve released it yet.

• Airbnb CEO Brian Chesky with an interesting take on interfaces in the age of AI and why chat isn’t a panacea: “I do not think a chatbot is the right interface for travel or e-commerce…I think the future is not apps. The future is agents, but I don’t think they’re going to be text-forward. I think they’re going to be really rich user interfaces…Imagine using iMessage to do everything, when in fact every other app has a unique interface…With e-commerce, you want a very rich user interface. It would be agentic. You can have a conversation with it, but the point is that it has to be more visual.”

• OpenAI employees were reportedly given the chance to cash out up to $30M each in secondary sales.

Every week I summarize the most interesting things happening in AI, in 5 minutes or less. Follow @tjhoyos on X/Twitter and subscribe to get this in your inbox every week!

Subscribe now

Subscribe now

Three big things

1. Prediction: you’re going to hear a lot about “company brains” or “company intelligences” in the coming years. We (and so many others) are working on building our own versions internally. The idea is to combine all the disparate data points your company has (product analytics, internal email and Slack threads, customer support emails, public reviews, ad performance, literally everything) into a single repository that an AI system can understand. Then people at the company can ask it questions to make better decisions, and eventually it starts being proactive by suggesting actions to take (or taking action itself) to maximize progress. There isn’t a go-to tool people use right now because the shape of every company is different, but stuff is starting to pop up: Scout is an open-source project in this area. Relatedly, the inimitable Ann Miura-Ko from Floodgate wrote a great piece on the different levels of AI-native companies.

2. Data show that models from Chinese AI labs trail their US counterparts by several months. It’s clear both that the frontier of AI is still in the US and that China is a fast follower and fierce competitor. And they are currently winning in open source, which is a problem for long term competitiveness.

3. OpenAI missed its internal 2025 targets for ChatGPT usage, including a goal to hit 1B weekly active users, and OpenAI CFO Sarah Friar told the company she’s worried about meeting compute spend commitments if they don’t grow faster. The stocks of some of their large public partners (Oracle) dropped on the news in recognition the AI labs are a bit a high-wire act with truly jaw-dropping capex contracts. A leading company slowing down is one of the only things that could send AI as a category into a tailspin. All that said, apparently both revenue and adoption of GPT-5.5 exploded in Q1 as they focused more effort on Codex, agents and enterprise, and they’re back on track, so I’m not worried about their prospects.

The rest

• AI was a major driver of US economic growth in the first quarter, with estimates that it accounted for 75% of all growth.

• YC’s advice on how to build a company natively around AI from day one: “AI as operating system, not tool…closed loops everywhere…make your company queryable…software factories…no more human middleware…three employee archetypes (per Jack Dorsey)…token-max, not headcount-max…the early-stage advantage.” They go into more detail and it’s worth a read.

• Interesting: the Founder of Khan Academy Sal Khan is launching a $10k AI degree with partners like Google, Microsoft, and Replit. There isn’t a legit AI credential and traditional college education feels a little less relevant every year, so I’m not surprised people are trying to innovate here.

• OpenAI is finally wising up to the fact that they have to paint a positive vision of society’s future with AI. There has been a noticeable shift in their public communications. They’re no longer harping on the prospect of job loss and claiming they will automate everything. They’re trying to respond to concerns about data centers (resource usage, electricity prices, appearence etc.) by showing the impact on real communities, e.g., this interview with the mayor of Abilene, TX, the location of their massive Stargate data center. I really hope they keep this going.

• Humanoid robots in the wild: “Japan Airlines will trial humanoid robots for baggage handling and aircraft cleaning at Tokyo’s Haneda Airport starting in May.”

• The future is getting weirder: OpenAI’s new GPT-5.5 model has a strange habit of talking about goblins, gremlins, and other creatures when using metaphors. Apparently this is because of the “nerdy” personality feature.

• Google reportedly controls ~25% of global compute capacity.

• Very cool: some researchers developed an LLM trained only on text from before 1931. The big idea here is to see how well these models can predict the future (given we know how things unfolded after 1931), and how we can apply that knowledge to making current models better at predicting the future from today.

• Lol: “My job used to be writing code. Now it’s marriage counseling for Codex and Claude Code.”

• AI is political and the teams are confusing: Senator Bernie Sanders is discussing AI on a panel with the dean from the Beijing Institute of AI Safety and Governance.

• Dan Shipper thinks the dominant UX of knowledge work will be a side panel with chat and agents on the left, and your work on the right, with you collaborating in real time.

• People are using AI to seek guidance. Anthropic published an interesting breakdown of what people asked for advice on across 40,000 Claude chats. The top use cases were health and wellness at 27% (interpreting test results & chronic conditions, injuries, respiratory symptoms & treatment, calories & macros for body composition), and professional and career at 26% (job search & opportunity identification, career transitions & path selection, offer evaluation & salary negotiation). The latter is extremely relevant to our company’s work.

• Context on Anthropic’s growth rate: “And the company previously known as OpenAI’s junior competitor has become possibly the fastest-growing business in the history of capitalism. Anthropic’s revenue is increasing faster—much faster—than Zoom’s during the pandemic, Google’s during the early 2000s, and even Standard Oil’s during the Gilded Age. If the company’s current growth rate were to continue, then by early next year it would be taking in more money than any company in the world.”

• Interesting Stratechery interview with Sam Altman. Some highlights: “pre-training and post-training will converge as a singular training stack for ai models (very bullish inference test-time compute)…model and ‘wrappers’ are basically the same thing (very good for cursor’s valuation)…openai may move to ‘pay-per task’ pricing vs. per token. token pricing is dead. GPT 5.5 is more expensive BUT USES FEWER tokens…HUGE win for amazon’s trainium chips that will now run chatgpt: Sam: ‘We are quite excited to get these models running on Trainium.... over time, more and more of it will be on Trainium.’…the new managed agent’s product = enterprise chatgpt agents for all of amazon’s clients. openai might just snatch back the enterprise market share they lost”

• GPT-5.5 Pro is probably the best model in the world right now, setting a new high on Epoch’s composite capability benchmark which tracks performance in a bunch of areas with one number.

• Menlo Research released an open-source humanoid robot platform.

• Ineffable Intelligence, a new AI lab from a former Google DeepMind employee, raised a $1B+ seed round.

• China is attempting to block Meta’s $2B acquisition of Manus, which was founded by Chinese citizens, but moved its HQ to Singapore.

• OpenAI launched a new real-time voice model.

• OpenAI’s GPT-5.5 might be more cost-efficient than Anthropic’s Opus-4.7: “GPT-5.5 is ~39% cheaper than Opus 4.7, across merged PRs bucketed by diff size…despite the higher output token cost, 5.5 is cheaper for input tokens (cache writes are free), more token efficient, and tokenizes the same text to fewer tokens.”

• DoorDash introduced a new “Tasks” app that lets gig workers take on additional side jobs beyond food delivery, and many of the tasks are AI or robotics adjacent: “Film household chores (dishes, laundry) for humanoid training data…Scan store shelves (monitor inventory / product placements)…Close Waymo doors left ajar”

• AI is increasingly becoming a national security and policy priority. Here is Treasury Secretary Bessent weighing in: “If we don’t win in AI, then it’s game over.”

• Honestly I’m kind of surprised by this collab: Google and Exa (an AI search company with a focus on agents) partnered to bring Exa’s search and grounding capabilities into Gemini models. I would have thought Google would just try to own this themselves.

• OpenAI and Microsoft updated their partnership.

• Google entered into an agreement with the US DoD that allows its AI systems to be used for government purposes.

• The explosion in AI-driven coding is putting strain on platforms like GitHub where usage charts are going vertical. It’s made spam and scale much harder to manage.

• xAI is reportedly training Grok 5 on an extremely large compute cluster, maybe the largest ever.

• Box CEO Aaron Levie on why AI increases demand for software developers: “software jobs aren’t going away. Agents are the single biggest form of leverage for anyone technical in history. Probably has never been a better time to be technical in terms of being able to accomplish something solo, in a team, or company.”

• Mercury launched a command-line interface that lets you manage banking and other financial tasks from the terminal. Most products will end up having CLIs.

• Parallel.ai, which believes there will be many more AI agents browsing the internet than humans, is building infrastructure and tools around that vision. They raised at a $2B valuation.

• In the marketplaces of the future, will people outsource buying and selling to agents? Anthropic ran an experiment where they did that and the agents successfully matched participants and made deals without any human negotiation.

• Rogo, who built one of the first AI investment banker products, raised $160M.

• Gemini can now generate files such as documents, spreadsheets, presentations, and PDFs directly within chat.

• The best models in the world get exponentially smarter as they “think” more. In a recent cyber-attack simulation, researchers had GPT-5.5 and Anthropic Mythos think harder and longer (i.e., consume more tokens before responding), and it wasn’t clear there was a ceiling to their performance. After 100 million tokens, performance was still going up (the chart in the link is amazing). This is extremely bullish for upcoming model generations.

• SoftBank is reportedly launching a robotics-focused company to automate the construction of data centers.

• AI automation company Gumloop introduced dedicated email addresses for AI agents into their product.

• Meta released an MCP and CLI to make it easier to port your ad spend data into AI tools like Claude and ChatGPT to analyze it, build reports, etc.

• This was thought provoking: “Unprecedented abundance will create unprecedented demand for things that cannot be made abundant. That might mean local things: really unique restaurants, theaters, hotels. It might mean new towns altogether. I have a hunch that drones, EVs, and EVTOLs [electric aircraft] should expand local frontiers. For an enterprising new settler, it might make sense to start gobbling up cheap, buildable land where it doesn’t currently make much sense, or open up new frontiers beyond the land. You might get it wrong, you probably will. No one said scarce was easy.”

• Series is an AI social network that lives inside iMessage and raised $10M.

• OpenAI launched an anthropomorphized character inside Codex called “Pets.” Not sure what they’re trying to do here.

• Ethan Mollick in the Atlantic: “we whipsawed from ‘AI is a bubble’ to ‘there are not enough data centers’ in less than six months. Spoiler: its agents”

• Citadel published analysis that frames AI as a net jobs engine rather than a force for displacement, especially in software engineering.

• Strong argument from NVIDIA CEO Jensen Huang on the dangers of doomerism and why AI is a positive force for the economy and labor markets: “On one hand, maybe a scientist thinks by warning people that AI is going to completely permeate and proliferate across radiology and therefore radiologists are gonna get wiped out…that might be considered warning and therefore helpful. But in fact it’s hurtful. If we convince everybody not to be radiologists and now we need radiologists, that actually is hurtful to society. It’s hurtful that we convince all the young college graduates not to study software engineering. Turns out the United States needs more software engineers than ever. That’s hurtful. So we have to be mindful of how we communicate the importance of this technology and what it’s able to do — to advocate for policy, advocate guardrails on one hand. And on the other hand, scaring people with things like saying nonsensical things which are not gonna happen. That this is an existential threat to humanity. There’s a 20% chance that it’s existential. That’s ridiculous. Now, that it’s going to wipe out 50 percent of new college grad jobs. That it’s going to completely destroy democracy. I mean these kinds of comments are not helpful. They’re made by, you know, people who are like me — CEOs. You become a CEO, maybe you adopt a God complex. Before you know it, you’ll know everything. So I think we have to be careful and really ground ourselves to talking about the facts. The fact is this: AI has created more than half a million jobs in the last couple of years. The fact is AI is our greatest, our best opportunity to reindustrialize the United States, to bring manufacturing jobs back to the United States. The fact is that’s gonna generate hundreds of thousands of jobs, trillions of dollars in new economy back into the United States. The fact of the matter is companies that use AI have demonstrated the ability to grow faster. When they grow faster, they hire more people. Apparently AI creates jobs.”

Every week I summarize the most interesting things happening in AI, in 5 minutes or less. Follow @tjhoyos on X/Twitter and subscribe to get this in your inbox every week!

Subscribe now

Subscribe now

Three big things

1. Meta will use monitoring software to record the screens of their employees to build a dataset they can use to: (1) train AI models, (2) analyze how people work day-to-day including identifying top performers and inefficient workflows, (3) eventually automate tasks or give feedback to humans doing them in real time. This might sound crazy, but I think we’ll see WAY more of this. Relatedly, OpenAI is testing a new Codex feature that lets it retain work context from your daily activity via screenshots. This is why a lot of people are worried AI will automate knowledge work, which largely happens on computers.

2. Huge news out of SpaceXAI and coding startup Cursor, the latter of which apparently had deeply negative gross margins at -23% as of earlier this year. The companies signed an agreement whereby they’d team up to train a new coding model using Cursor’s coding dataset and SpaceXAI’s massive data centers, while also giving SpaceXAI the right to buy the coding company for $60B when the experiment is over. They’re aiming to compete with OpenAI and Anthropic.

3. Dylan Patel from SemiAnalysis discussed the negative public sentiment around AI: “people hate AI. AI is less popular than ICE. Less popular than politicians…[advancements will cause] more downstream business problems. People are going to get more and more scared of AI. They’ll start blaming more and more of their own problems [and global problems] on AI…you look at the comments on the news articles where Sam Altman had a molotov cocktail thrown at his house twice in two weeks — people are cheering it on. This is just the beginning. We’ll see large scale protests against AI in 3 months.”

The rest

• AI’s impact on labor markets is unclear with various conflicting data and narratives:

  • Nvidia CEO Jensen Huang: “The narrative of AI destroying jobs is not going to help America — It’s false…Somebody said that AI is going to destroy all of the software engineering jobs…We now have agentic AI inside Nvidia…The software engineers are busier than ever.”

  • Public companies are showing signs of contraction in white-collar jobs, with total S&P 500 headcount declining after years of steady expansion.

  • Meta will cut around 10% of its workforce, roughly 8k roles. They join a string of recent companies (Block, Atlassian, Oracle) with large layoffs.

  • New analysis out of the UK: no clear evidence that AI has reduced overall employment. Roles with higher exposure to AI have grown faster than less-exposed ones. However, AI-exposed jobs have seen weaker pay growth since 2019, though this began before generative AI.

  • Anthony Pompliano on why he changed his mind and now believes AI will lead to job growth: “The number of software engineers being hired has been increasing. The number of open software engineer roles is growing. The number of new college grads who get hired has increased 5.6% over the last 12 months. The unemployment level for people aged 20-24 years old who have a college degree has fallen from nearly 9% to almost 5% as well…If AI can make employees more productive, which is widely accepted as fact, then companies are going to want as many productive units of labor as possible. This is a key reason why I am changing my mind. AI appears to be a magical technology that will make companies more productive and more profitable. The net result will be more corporations, more startups, and more jobs.”

  • Interesting argument for why AGI doesn’t lead to mass unemployment, as there will always be demand for what is fundamentally human: “the durable jobs of the future won’t be about monitoring AI systems or prompt engineering. Those are transitional roles in the automated sector. The durable jobs will be in the relational sector, where the human element is the product itself. Some already exist and are growing: nurses, therapists, teachers, boutique fitness instructors, personal chefs, bespoke tailors, craft brewers, live performers, spiritual guides, childcare workers, and many varieties of hospitality and care work. Others are emerging: experience designers, human-AI collaboration artists, provenance certifiers, community curators. Many haven’t been invented yet, just as six out of ten jobs people hold today didn’t exist in 1940. The most common pushback I get when I say this is: “but not everyone is creative, not everyone will be artists.” I think this misunderstands what’s being asked. You don’t need to be Picasso. You need to be the person whose involvement makes the product feel like it was made for someone, by someone. The economics of structural change tells us that when technology makes one type of production cheap, the economy doesn’t collapse. It transforms. It shifts toward the things that technology can’t make cheap. For AI, those things are exactly the ones where human involvement carries inherent, irreplaceable value.”

  • Box CEO Aaron Levie on why humans are necessary to the adoption of AI inside companies: “The real world will need a ton of help [from people and consultants] actually getting agents going in the enterprise. Companies have legacy tech stacks they need to modernize, data in tons of fragmented tools, knowledge that isn’t captured or digitized, and change management needed to actually utilize agents effectively. And they have to do all this while still running their business day-to-day, unlike startups. This is why there is so much opportunity for companies (software or services) to actually deploy agents in specific domains and workflows. This remains a big opportunity for both existing services providers but also tons of new startups as well. Every new technology wave produces a new era of consulting firms that can deliver on that technology. It’s also why the FDE model is going to be alive and well for a long time because companies will want to have their vendor actually help drive the change management and implementation for their new workflows. The people aren’t going away. Far from it.”

• Marc Andreessen on why AI-driven automation is a good thing: “The book Bullshit Jobs by David Graeber was wrong in substance but right in gestalt. Many jobs of the past would horrify us today -- chimneysweep?? Many jobs of today will horrify our descendents. We actually made people DO that? For their whole LIVES?” He’s not wrong! I think there’s a forward-looking element to this as well: someone from 1910 would have laughed at the idea that an influencer recording selfie videos would be considered “work.”

• OpenAI launched GPT-5.5, their latest flagship model with particularly strong capabilities in agentic tasks, while also increasing efficiency.

• Demand for tokens is increasing exponentially.

• AI is increasing productivity according to economic indicators from Morgan Stanley: output per worker is accelerating after a period of stagnation.

• Open source models rely on distillation, i.e., using the outputs of frontier models as data to train a smaller model. The evidence in the link above is that open source models tend to be extremely similar to either ChatGPT or Claude in their responses, indicating the preferred source data.

• Article on how US AI data center expansion is being slowed by infrastructure bottlenecks like power and transformers: “Of 140 planned US datacenter projects…only about 5 gigawatts are actually under construction. A significant share of the physical infrastructure enabling one of America’s largest infrastructure bets is being sourced from China.”

• Dylan Patel argued that Anthropic’s meteoric rise in model quality and revenue don’t mean that OpenAI is losing. He thinks demand dynamics (and ability to serve demand with access to compute) matter more than rankings, and on that front OpenAI may actually be ahead: “There is such demand for these tokens and such limitations on compute… it’s pretty clear even the tier 2 or tier 3 labs are going to be sold out of tokens [with] infinite demand for intelligence.”

• Relatedly, given it’s compute-constrained, it’s no surprise Anthropic (1) expanded its partnership with Amazon in a big compute deal: 5 gigawatts of additional data center capacity, $100B+ in long-term cloud spend, continued use of Amazon’s Trainium AI chips, and more. (2) took in another multibillion-dollar investment from Google.

• OpenAI, Sam Altman and Elon Musk are set to face off in court in a high-stakes dispute over whether Elon is entitled to ownership in OpenAI, and whether it was fine for the company to change from a nonprofit into a for-profit entity.

• A new open-source, coding-focused model from Chinese AI lab Kimi is awesome.

• Very cool: Ontora is a new tool that uses AI to interview employees to learn about bottlenecks, inefficiencies, and areas of improvement. This is an early example of how AI will shape organizations as it’s increasingly applied to traditional management and operations techniques.

• Cowork now lets Claude generate “live artifacts” like dashboards and trackers that stay connected to your apps and files, updating automatically with fresh data.

• Talktoreview is a new tool that uses AI to get reviews from customers with shareable links, QR codes, and conversational experiences.

• Meta with some smart PR (and hopefully positive impact on the electrician shortage)! They launched a short, free training program to help people break into entry-level fiber technician roles in data centers.

• Interesting quotes from Google AI chief Demis Hassabis in a wide-ranging interview: “Not all countries are pessimistic about AI. I was just in India for the AI Summit Modi had and they’re quite optimistic there…The most exciting current prospect in AI is our work at Isomorphic Labs. AlphaFold is just one of the many problems we need to solve. We need 6 ‘AlphaFold’ moments to compress the drug delivery timeline from 10yrs to a few months…I think of a post-scarcity world where on the bright side we will have an unbelievable amount of science but we will have to think of economic problems of sharing proceeds equitably. We will also have philosophical questions to answer and need great new philosophers”

• A new paper from Yann LeCun’s team looks like they’ve made great progress on world models as an alternate path to AGI besides LLMs, but some of the public conversation is overstated. The results of the work is more efficient systems that can learn physical structure and perform certain tasks faster than larger approaches. It calls into question the dominant view that “bigger is better” and suggests a possible future where models learn compressed representations of reality rather than predicting raw outputs.

• There are two new “Deep Research” modes in the Gemini API: fast/efficient or extended computation for higher-quality analysis. It’s kind of crazy that we can just tell the models to “think longer and harder” and we get better answers.

• OpenAI coding tool Codex adoption is accelerating. They just hit 4M weekly active users.

• AI music app Suno hit #1 in the app store.

• There is a new OpenAI image model and it is amazing. Some of the examples are awesome.

• Former OpenAI researcher Jerry Tworek launched a new AI lab called Core Automation.

• A growing view in AI research is that model scale is compensating for poor data quality: “[Andrej] Karpathy told Dwarkesh [Patel] that a 1 billion parameter model, trained on clean data, could hit the intelligence of today’s 1.8 trillion parameter frontier.

  That is a 1,800x compression claim.”

• Anthropic’s Mythos model, despite being deemed too sensitive for broad release, was reportedly accessed by unauthorized users due to a series of operational security gaps.

• Google is open-sourcing a standard for helping AIs build products that use a specific design system: the DESIGN.md file. We got very excited when we saw this internally as it would make it easier for our coding agents to build stuff that fits in with our existing design system.

• OpenAI is rolling out “workspace agents” in ChatGPT: shared agents that run workflows across teams and tools.

• In ~two years, Google went from having basically no AI-written code to around ~75% of its codebase being generated by AI.

• A new prototype called “Flipbook” is very weird and interesting. It’s a kind of design/generation tool where the screen is generated entirely by a model, streaming live pixels instead of relying on HTML, layouts, or rendering engines.

• Intercom CEO Eoghan McCabe is re-branding their AI customer service agent to be more general, e.g., their flagship product Fin now supports other roles like a Sales Agent. They’re trying to build a unified customer agent spanning the entire lifecycle.

• OpenAI is launching a clinician-focused version of ChatGPT.

• OpenAI released an open-weight model for redacting personally identifiable information (PII) from text.

• One founder’s view that specialized AI memory systems are going away and the world will move toward search-based retrieval.

• A new product called Clicky lives on your screen and points to and shows you what to do using AI. Very interesting UX pattern.

• Researchers found that AI-based hiring systems tend to favor resumes that have been written by AI over those written by humans, even when qualifications are identical. Models show a strong self-reinforcing bias. This puts you at a huge disadvantage if you’re 100% hand-writing your resume and companies are using models to screen candidates.

• A study calculates that the big AI tools (ChatGPT etc.) are creating significant consumer welfare gains, measured through how much users would pay to avoid losing access.

• AI is getting ever more political. Bernie Sanders weighed in on the existential risk of AI. Here is Melania Trump: “We are not here to prepare our children for yesterday’s world. Be purposeful with your objectives, and remember that AI accelerates everything.”

• Anthropic ran an experiment where they built an internal goods marketplace for their employees where AI agents acted as negotiators. The models successfully completed hundreds of transactions (e.g., it bought one of their employees a snowboard) worth thousands of dollars. People generally viewed the outcomes as fair.

• AI will be an even bigger player in financial markets. A team gave Claude Opus 4.6 a small trading balance and the model grew it dramatically on a prediction market.

• Relatedly, Instacart co-founder Apoorva Mehta launched a hedge fund that uses AI agent swarms to conduct research and make trading decisions. Abundance says their mission is to improve capital allocation.

• Aaron Levie on why AI is increasing work instead of reducing it.

• Why AI supercharges some workers: “everyone assumed ai would flatten the talent distribution.. turns out it amplifies the hell out of it. it used to be: can you build it. now it’s: do you know what’s worth building, & can you feel when it’s wrong. that’s ~unteachable & ~unautomatable right now. models can generate 100 variants of anything but they still can’t tell you which one matters. amazing talent is roughly priceless in the ai era because with ai it’s leverage++++++.”

Every week I summarize the most interesting things happening in AI, in 5 minutes or less. Follow @tjhoyos on X/Twitter and subscribe to get this in your inbox every week!

Subscribe now

Subscribe now

Three big things

1. AI’s impact on labor markets and layoffs:

   1. Snap is laying off a significant portion of its workforce, joining a string of leading tech companies (Block, Atlassian, Meta, Oracle). Many think this is being driven by AI reducing headcount needs.

   2. This awesome podcast with Block CEO Jack Dorsey goes into extreme detail on how they’re reshaping the company with AI. Strongly recommend a listen.

   3. Marc Andreessen on whether AI is driving these recent layoffs: “What you have happening right now is you have essentially every large company that is overstaffed…They all have the silver bullet excuse, right? Ah, it’s AI…Maybe the skill set of a lot of college graduates over the last decade doesn’t necessarily match to the job market. That’s a very uncomfortable conversation. But if you talk to any employer, they’ll immediately tell you that.”

   4. Elon Musk on a potential new social and economic contract: “Universal HIGH INCOME via checks issued by the Federal government is the best way to deal with unemployment caused by AI. AI/robotics will produce goods & services far in excess of the increase in the money supply, so there will not be inflation.” You know things are getting wonky when the richest person on earth is suggesting policies that are a far cry from traditional free-market capitalism.

   5. This survey shows many employees are resisting or undermining AI adoption inside their companies, possibly because they think it’ll hurt job security.

   6. Interesting paper on the game theory of AI automation. The authors think individual firms are incentivized to replace human workers with AI, which hurts consumer demand and economic output in the aggregate. For what it’s worth, I don’t think this is super plausible. While there will be turmoil in labor markets, AI will create many more opportunities than it destroys, and I just don’t see us consuming less as a society.

2. Shopify CEO Tobi Lütke shares stats on the increasing prevalence of entrepreneurship, and suggests it’s one of the most viable career paths in the AI-driven economy: “The most future-proof job: Entrepreneurship”

3. Open source is catching up. AI automation startup Lindy is seeing comparable evals performance with open models at half the cost of leading closed source models. He’s using GLM-5.1.

The rest

• Affirm CEO Max Levchin with an interesting point: “I think the average IQ, with AI in your ear at all times, is about to go up to 150—which is north of the ‘genius’ definition.” How does the world change when you and everyone you interact with is effectively a cyborg with AI wearables?

• Sign of the times (and the future getting weirder): a once-hot consumer shoe brand Allbirds pivoted away from apparel into AI infrastructure! I do not have high hopes here…

• Anthropic is pushing back against proposed legislation in Illinois that would limit legal accountability for AI companies in cases involving severe harm or large-scale damages caused by their models. In some ways, this feels like the social media arc all over again.

• Sentiment in the US around AI is surprisingly negative.

• Salesforce is evolving from a company that builds software to one that serves humans and agents. They’re shifting the platform to an API-driven, headless architecture. Developers (and agents) can access core products like Salesforce, Slack, and Agentforce via APIs, CLI, and other agent-native interfaces.

• OpenMythos is an open-source research project that attempted to recreate Anthropic’s unreleased Mythos model. They used a similar architecture to the one Anthropic allegedly used (looped transformer with a mixture-of-experts router) and they saw impressive results. This may be one of the architecture breakthroughs the industry has been looking for.

• The case for ChatGPT apps made by A16Z’s Olivia Moore: “ChatGPT could be the first true global super app. It holds both context on you and the payment layer…Imagine you’re driving and can say ‘order my usual at the nearest Starbucks’ and it executes…There are many apps we’d never have to open again”

• Two new interesting stories around data: (1) An emerging human data marketplace reached a major milestone with over a billion user-contributed data uploads and fast growth in structured, labeled datasets. I wonder what the market for data will look like. Will I get paid for my data some day? How differentiated does it have to be to generate income? (2) A new human-in-the-loop service is emerging for AI agents, where they can escalate difficult tasks to verified subject-matter experts in real time.

• Sam Lessin thinks the most profound impact of AI is not economic displacement or safety, but the impact on meaning and purpose. We agree and it’s a big reason we’re building the company.

• AI might help undo some of the polarization caused by social media: “Social media is populist and polarising. AI may be the opposite.”

• Data showing a clear divide in how different groups perceive AI: “Men aged 18–54 are much more partial toward AI than other groups (+32 net favorability in Dec 2025).”

• Relatedly, the demographic patterns of early ChatGPT usage skewed male, but that has balanced out over time.

• OpenAI is evolving its ChatGPT ads product toward more performance-oriented models like Instagram and Facebook. This makes it much more likely we, for example, would try using them!

• There’s a new version of Claude Code that’s heavily geared towards running multiple agents at once. It’s a signal of their top priorities and how ppl are using the product…

• AI HR company Diana pivoted to OpenClaw for businesses. Curious to see how this does! I know a ton of companies are using OpenClaw internally, but there are serious security tradeoffs to using that set up “out of the box.”

• From Kobeissi on the latest AI bottleneck: “The US AI boom has a hardware problem…~50% of all US data centers planned for 2026 are expected to be delayed or canceled because [electrical equipment like transformers, switchgear, and batteries are] in shortage.”

• LinkedIn is expanding into the AI data procurement market. They’re planning to generate labeled data and evals that AI labs can use to train their models. They’ll compete with companies like Scale, Surge, Mercor, Handshake, Micro1 and more.

• Skye is an interesting new AI consumer play where they’re building an AI-powered widget that will live on your iPhone homescreen and give ambient, context-aware help instead of relying on the user to click into specific apps via the traditional home screen grid.

• China enacted a regulatory framework governing anthropomorphic AI systems (i.e., human-like chatbots, avatars, deepfakes, etc). China has the capacity (and generally tends to) be much more heavy handed than places like the US when it comes to interventions that they think will promote objectives like stability and safety. I’m interested to see how this plays out.

• OpenAI’s CRO sent a policy memo to their team emphasizing focus on core products (including enterprise) and building a moat, while also taking a shot at Anthropic by accusing them of inflating their revenue numbers.

• AI is influencing politics and elections: “ai predictors of humans are being used to make targeted, persuasive [political] ads…ai agents are calling people to get them to vote some way…elected officials are posting deepfakes.” This is only the beginning.

• Google released a new version of their robotics-focused AI, Gemini Robotics ER 1.6.

• Hedge fund Jane Street Capital is looking to invest $1B in AI infrastructure company Fluidstack.

• In an awesome episode of the Dwarkesh Patel podcast (strongly recommend a listen), NVIDIA CEO Jensen Huang argued that despite the rise of alternative AI chips like TPU (Google), Trainium (Amazon), and other custom ASICs, NVIDIA will remain dominant. My sense is he was hinting at the word “monopoly” but couldn’t quite come out and say it directly. There was also an extremely interesting conversation around export controls and China.

• A humanoid robot from a Chinese tech company set a half marathon time record in an event designed for robotic runners, before crashing right before the finish line…crazy video and story. Between this and the fact that robots are being used on the frontlines of the war in Ukraine, I think we’re seeing the early innings of a revolution in humanoid robots.

• Traditional hiring processes are breaking down under the pressure of AI. We need a better way (and we’re building one).

• An emerging thesis is that AI’s biggest opportunity is delivering and selling end-to-end work outcomes (not software). This changes AI from a productivity layer into a labor substitute, and increases the market size dramatically.

• Eigen teased a new AI social product and announced a fundraise from Benchmark.

• Anthropic released Claude Opus 4.7, their newest generation model.

• The latest Codex update expands it from a coding assistant into more of a development environment agent with developer tools, external services, and system-level workflows.

  Every week I summarize the most interesting things happening in AI, in 5 minutes or less. Follow @tjhoyos on X/Twitter and subscribe to get this in your inbox every week!

  Subscribe now

Subscribe now

Three big things

1. This is deeply unsettling: OpenAI CEO Sam Altman’s house was attacked twice this weekend, once with a molotov cocktail and later with gunfire. Negative public sentiment about AI (polls show 80%+ are “very” or “somewhat concerned”) and data center boycotts have now escalated to violence. This is obviously horrifying and those people are unhinged, and it might also be a wake up call: I strongly agree with Will Manidis’ argument that AI labs need to paint a better, clearer vision of the positive future enabled by AI, and that future should include real sacrifices from AI labs that allow the rest of society to participate in the wealth that’s created. Otherwise, I think we’ll continue to see terrible stories like this. This will be the defining political and societal issue of our time. I hope we’re ready for it.

2. The bull case for the AI-driven economy, from Marc Andreessen, “The ‘AI job loss’ narratives are all fake. AI = massive ramp in productivity = massive ramp in demand = massive jobs boom.” And he points to some early data: “Tech job openings rebounded sharply in 2026, challenging popular narrative that AI is wiping out engineering roles...more than 67,000 software engineering job openings, highest level in 3 years. Listings have doubled since a trough in mid-2023.”

3. Guru Chahal calculates the market opportunity for AI adoption in knowledge work: “Did some back-of-envelope math on AI market penetration vs. knowledge worker labor spend. The entire AI industry has captured <1% of a $6T+ market. Even the best verticals - software dev & customer service - are in the low single digits. Most are under 0.5%. And that doesn’t even fully take into account that these markets are expanding, not fixed!”

   

The rest

• Labor market signals suggest AI’s impact on hiring is uneven across age groups. Entry-level workers (e.g., 22-25-year-olds) are seeing noticeable employment declines with older workers in the same jobs seeing stability or growth.

• Thought experiment from Michael Dell: “PC & smartphone demand was always bounded by one thing: people. But what if billions (or more) of agentic AIs aren’t bounded by people at all? What happens when they never sleep…when they work, talk, design, buy, sell, and decide 24/7? What then happens to compute demand?”

• Relatedly, Cognition CEO Scott Wu thinks compute supply has grown at 3x per year for the last few years while AI demand (i.e., inference) has grown at 10x. According to him, if we don’t want to see prices increase, we should direct more traffic to smaller, specialized models.

• Meta AI is back. They launched their new flagship model Muse Spark and it’s quite strong. It’s not state-of-the-art, but they’re within striking distance. Their mobile app charted in the App Store and the stock was up +10% on the news.

• Perplexity is also back. Their “Computer” agent product is driving much faster adoption and their ARR surged from ~$300M to over $450M in a month.

• However, they both have a long way to go to catch up with Anthropic, who is experiencing hypergrowth: ARR increased from $9B to $30B in 3 months, they added $11B in March alone, and in a single month they generated as much revenue as Palantir, Anduril, and Databricks combined.

• According to Brad Gerstner, AI lab margins are improving quickly as they benefit from existing compute build-outs and revenue scales rapidly. He says it’s one of the fastest increases in gross margins he’s seen in memory.

• Compute is a central battleground in AI, and the distribution is concentrated. Data by Epoch AI shows that the world’s AI processing power is overwhelmingly controlled by US hyperscalers, with Google leading the pack. Chinese firms hold just over 5% of global AI compute and their share is shrinking under export restrictions.

• Poke is a consumer AI product bringing agent-style AI to mainstream audiences through iMessage. Apparently their retention is off the charts.

• Ronan Farrow from the New Yorker wrote an exposé (hit piece?) on Sam Altman. It doesn’t sound great, but hard to separate fact from fiction here.

• Leaked financial data shows OpenAI and Anthropic seeing crazy revenue growth ahead of potential IPOs, but also soaring model training and compute costs, e.g., OpenAI expects $121B in compute spending by 2028 with a $85B loss in that year alone.

• Speaking of IPOs, apparently OpenAI CFO thinks they’re not ready for primetime and they’ll face headwinds in public markets. My take: I really, really hope all the labs go public so I can buy the stocks.

• Anthropic is seeing engineering output explode thanks to AI, but it has not obviated the need for product management and design. The company is both hiring more PMs and deputizing engineers as “mini-PMs.”

• Sam Altman: two breakthroughs we need on the way to AGI are continuous learning and persistent memory.

• Ben Thompson hates OpenAI’s acquisition of tech media startup TBPN.

• NVIDIA launched an extremely impressive new voice model with ultra-low latency.

• The bull case for Google: “Probably only company that controls the full stack. chips + data centers + models + distribution (search, android, youtube, workspace).”

• Marc Andreessen on how much state-of-the-art AI experiences cost right now: “Magical OpenClaw experiences that use frontier models cost $300-1,000/day today, heading to $10,000/day and more. The future shape of the entire technology industry will be how to drive that to $20/month.”

• And here he is again with more context on supply and demand: “Frontier models become more expensive to serve. Token demand explodes at exponential rates. Solve for the equilibrium.”

• Clicky is an AI-powered learning companion that sits alongside your cursor. Interesting launch from Farza!

• A paper on “machine culture” discusses the risk of AI creating a monoculture. But Mickey Friedman points out that even after AI became dominant at human games (Go, Chess), the capabilities of human players then got much better and created net new techniques.

• HeyGen has a new AI avatar model, Avatar V. These are getting scary good…

• Applied Compute raised $80M.

• AI BDR company Artisan raised $35M and launched their new sales agent, Ava.

• Nic Carter’s opinion on AI: “It should be pretty obvious at this point that AI is a ‘force multiplier’ not a ‘labor substitute’. It helps experts be better at things they are already good at. It doesn’t let beginners match experts. If you can’t write, anything you write with AI will be unmitigated slop. If you aren’t a software engineer, anything you vibecode with AI will have security holes and won’t be able to scale past a toy demo. If you blindly trust AI to deliver on a research task without knowing the subject matter, you won’t be able to fact-check it. There’s this weird misconception of AI as something that completely levels the playing field. I don’t see it that way at all. There are mathematicians deriving novel lemmas with off-the-shelf models. Normal people can’t do that. AI is a tool that makes experts better. It doesn’t make everyone into an expert.”

• Claude Managed Agents introduces a full-stack solution for building and deploying AI agents.

• Anthropic is introducing an “advisor” architecture on the Claude Platform. How it works: a lower-cost model will consult a more capable model when it needs to, increasing quality and token efficiency while reducing costs.

• AI agents can now open a business bank account via Treasury startup Meow. This is only the beginning…

• Avec raised $8.4M to build an AI-enabled email client.

• Chinese open-source AI models are increasingly powering parts of Silicon Valley’s stack (Cursor, Cognition, Shopify, Airbnb).

• Semianalysis’ Dylan Patel thinks power is less of a constraint to scaling AI than people assume, and he’s more worried about memory and logic.

• One of the top-earning YC startups uses AI to match women for friendships.

• Claude for Word is bringing their AI to one of the biggest productivity apps on the planet.

• Bonobos founder Andy Dunn is launching a new AI app called Pie. It seems like they want to connect you to friends and events (or encourage you to reconnect with people you’re already close to).

Every week I summarize the most interesting things happening in AI, in 5 minutes or less. Follow @tjhoyos on X/Twitter and subscribe to get this in your inbox every week!

Subscribe now