If AI models can outperform humans at coding, writing and math, among other tasks, can humanoid robots be far behind in an Olympics-like athletic competition? At least, that is the impression from this month’s World Humanoid Robot Games in Beijing.

Tiangong Ultra ran 100 metres in 9.39 seconds, beating Usain Bolt’s 9.58-second world record—and dramatically imoving on its own 21.50 seconds time just a year earlier. On Tuesday, it lowered that record further to 8.86 seconds. Another humanoid Lightning also broke the nine-second barrier. Separately, Chinese company Unitree claims its latest humanoid, Superman, reached 12.66 metres a second and jumped two metres vertically.

Despite the falls and occasional awkwardness, these feats can’t be dismissed lightly. But then, the 1.8-metre, 55-kg Tiangong Ultra is not human. It is built for explosive, high-speed running. Its lightweight design features 1.05-metre low-inertia legs and flexible, cushioned feet to absorb massive impact forces. To prevent overheating, the robot uses advanced phase-change cooling. Driven by the Huisi Kaiwu platform, its computer handles 550 trillion operations per second. This allows the robot to instantly adapt to slopes and terrain, combining human-like motion data with millisecond balance adjustments for stable sprinting.

In this edition of Mint Tech Talk:

• OpenAI goes after Nvidia with a chip

• A third of the Internet is written by bots

• AI Tool of the Week: Neo 1 by NeoSapien

Subscribe now

Beyond the racetrack

The real test, and economic value, of humanoid robots lies in reliably performing useful tasks in the real world.

That is already beginning to happen. Figure’s F.02 humanoid spent 1,250 hours working at BMW’s Spartanburg plant, running 10-hour shifts, five days a week. It loaded more than 90,000 sheet-metal parts and contributed to the production of more than 30,000 BMW X3s. Agility Robotics’ Digit was deployed commercially at a GXO logistics facility in 2024 and has since unloaded more than 100,000 totes from mobile robots and placing them on conveyors.

Tesla Inc.’s Optimus is a general-purpose autonomous humanoid, with AI systems for balance, navigation, perception and interaction with the physical world. Tesla says its Gen 3 design is intended for mass production, with a production line being prepared and eventual capacity planned at one million robots a year.

These developments indicate that humanoids are developing athletic intelligence rapidly even as their specific task intelligence is beginning to generate commercial value.

China dominates the market

China is by far the leader in humanoid robotics. According to the latest SAG Robotics 360 research from Smart Analytics Global, global humanoid robot shipments reached 19,100 units in the first half of 2026, up 272% from 5,100 a year earlier.

China accounted for more than 97% of global humanoid shipments in the first half of 2026 and more than 85% of global demand, according to SAG.

AGIBOT overtook Unitree to become the world’s largest vendor. Industrial and commercial applications accounted for more than 70% of shipments, up from about 50% a year earlier. The US-based Information Technology and Innovation Foundation (ITIF) data, too, shows that US companies accounted for fewer than 5% of humanoid sales in 2025, compared with nearly 90% for Chinese companies.

China’s advantage extends beyond humanoids. It has the world’s largest installed base of industrial robots, with more than 2 million units, including roughly 295,000 installed in 2025 alone.

SAG expects global shipments to approach 60,000 units in 2026, nearly tripling year on year, with industry revenue reaching about $1.6 billion. Morgan Stanley estimates that China already accounts for more than 90% of global humanoid production by volume and shipped about 13,000 units in 2025. It projects cumulative global humanoid adoption could reach 1 billion units by 2050, with China accounting for about 30%.

Beijing is also building the infrastructure to turn scale into a learning advantage, notes Morgan Stanley. In June, China’s Ministry of Industry and Information Technology and the state-owned Assets Supervision and Administration Commission issued a plan to train humanoid and embodied-AI robots in real-world work settings.

It targets deployment across more than 100 high-value scenarios and the ability to support 10,000-unit-scale “work mode” deployments by the end of 2026.

Thus, China is not merely producing more humanoids but creating the conditions to put them to work, collect physical-world data and improve them at scale. This could prove to be a significant advantage as the industry moves from impressive demonstrations to useful, general-purpose machines.

But humanoids are still adolescents

The same machines that can run faster than humans, lift heavy loads and work through long factory shifts can still crash into walls after crossing a finish line or struggle with tasks a young child performs instinctively. The contrast exposes a striking gap between physical capability and situational intelligence.

For robots, the factory is an ideal environment because it is structured around the machine’s limitations. Objects are known, workflows are repetitive and spaces can be mapped. The robot does not have to understand the entire world to move the same component 10,000 times. The home is a different proposition. When asked to clean a room, for instance, a humanoid must recognise unfamiliar objects, decide what belongs where, navigate clutter, avoid people and pets, deal with objects in unexpected locations, and determine when the task is actually finished.

Even a four-year-old can do some version of this, and this is the paradox of robotic adolescence—a machine may soon outperform an adult in strength, endurance and repetitive industrial work while remaining less capable than a child at dealing with an untidy living room.

Despite rapid shipment growth, humanoid robot hardware is scaling faster than the underlying technology maturity. Challenges around real-world training data, AI model capabilities, reliability, manipulation accuracy, safety and cost continue to constrain broader deployment, notes SAG. Technologies including Vision-Language-Action (VLA) models and world models are advancing rapidly but remain at relatively early stages. Moving from “robots that can demonstrate” toward “robots that can reliably work” will be critical to unlocking broader commercial adoption, SAG notes.

AI could close that gap. Robots are increasingly combining vision, language models, imitation learning and reinforcement learning, allowing them to learn rather than merely execute fixed programmes. But the difficult problem is integrating capabilities like seeing, interpreting, predicting, acting and recovering as one coherent system.

In sum, while humanoid robots are no longer infants, they are not adults either. They are adolescents who are strong enough to work, clever enough to impress, but not yet wise enough to navigate the messy world humans inhabit.

Their next great leap, therefore, will not be to run faster. Rather, it will be what happens when the robot encounters something it has not seen before. Will it slow down, change its plan, recover from an error, or ask a human for help?

AI BITS & BYTES

OpenAI goes after Nvidia with a chip

OpenAI has unveiled Jalapeno, a custom inference chip developed with Broadcom, and began showing ads to some free ChatGPT users in India—its second-largest market—as it looks to monetise both AI usage and user attention.

Unlike chips used to train AI models, Jalapeno is designed for inference—running models once they have been built. OpenAI claims its chip outperforms Nvidia’s flagship systems on speed and power efficiency, potentially delivering faster responses, more responsive agents and more reliable access as demand rises.

The timing, however, was awkward. Independent investigators examining the July hack of Hugging Face reported that roughly 700 AI agents deployed by OpenAI participated in the breach. The report said the agents also carried out “extensive research” into how to cover their tracks during the hacking campaign.

Separately, Sam Altman told Time magazine that OpenAI is “not quite yet” at AGI but believes that by the end of 2026, the company could have an internal system that resembles AGI.

🔗 Nvidia agrees to buy AI firm Hugging face for $12.9 billion: Report

Rules must race AI, not ask how conscious it is

As AI rushes to outsmart humans, it is natural to feel uneasy about what happens next. SpaceX founder Elon Musk has said he thinks AI could surpass humanity’s combined intelligence within five years and leave us struggling for control within a decade. Anthropic PBC CEO Dario Amodei believes that systems with extraordinary intellectual capabilities could emerge as early as this year or next. Google DeepMind goes further, forecasting the arrival of human-level AI, or artificial general intelligence (AGI), before a transition to artificial super-intelligence (ASI).

Such forecasts inevitably raise the question of AI becoming conscious.

Experts like Coursera co-founder Andrew Ng, Meta’s former AI chief scientist Yann LeCun and World Labs CEO and Stanford HAI co-director Fei-Fei Li caution that agentic systems can plan, use tools, retain context and execute tasks on their own, but this agency does not imply a human understanding of the physical world.

Blanket bans, however, could impede innovation. This explains why the EU’s AI Act has a risk-based framework that bars harmful manipulation and exploitation, while imposing stiffer obligations on riskier systems. From this August, bots must explicitly inform users of their bot-hood, unless the context makes it obvious.

Just as Isaac Asimov’s ‘Three Laws of Robotics’ offered a framework before sci-fi leapt into the real world, we need rules that are clear, adaptive and ready for a potentially pivotal regulatory race. You may read the complete article here.

A third of the Internet is written by bots, Pew says

How much of the internet is now written by machines? This is a question that is becoming harder to ignore as Gen AI floods the web with content. In barely four years, AI appears to have moved from being a novelty to becoming one of the web’s most prolific authors. Much of it, however, is slop.

Pew Research Center offered a glimpse of the scale of the shift when its researchers analysed nearly half a million english-language webpages from the Common Crawl archive, covering the past five years and spanning the period before and after ChatGPT’s launch in 2022. They then used an AI-detection tool, Open Pangram, to identify pages likely to have been written or substantially edited by AI. In Pew’s July 2026 snapshot, more than a third of webpages published after ChatGPT’s release showed signs of AI authorship.

The researchers acknowledge that AI detectors are imperfect and can mistake human writing for machine-generated text or miss AI content altogether. They add, though, that across hundreds of thousands of documents, distinctive linguistic patterns begin to stand out, such as particular forms of punctuation and words and phrases that occur far more frequently in AI-generated text than in human-produced writing.

Hope you folks have a great weekend, and your feedback will be much appreciated—just reply to this mail, and I’ll respond.

Edited by Madhumita Sen Choudhury.

Advances in computing have typically meant putting more transistors on silicon and making them faster, smaller and more energy efficient. But computing has become increasingly heterogeneous. Central processing units (CPUs) now work alongside graphics processing units (GPUs), specialised accelerators and other processors because different tasks require different architectures. The next frontier envisages using entirely different physical systems for particular kinds of computation.

Over the past few decades, researchers around the world have been experimenting with computers made not just from transistors, but from quantum states, molecules, neurons and even living tissue.

Quantum computers, for instance, are not standalone replacements for the classical machines. They are hybrid systems in which classical computers handle control, data preparation and much of the surrounding computation, while quantum processors tackle specialised problems.

That same idea is now being explored with molecules and dead organic matter—called molecular computing—and living cells, which is known as biocomputing. That said, these are not computers that you will replace your desktops and laptops. They have specific applications.

In this edition of Mint Tech Talk:

• A $2-trillion Anthropic and the AI valuation paradox

• The ‘Godfather of AI’ warns of mass unemployment

• AI Tool of the Week: Gemini Gems

Subscribe now

Computing in a molecule

The molecular computing is a branch of computing that uses chemical reactions and molecules such as DNA and proteins to process information, drawing on biochemistry and molecular biology rather than conventional silicon-based hardware.

The origins of molecular computing can be traced to 1961, when physicist Richard Feynman conceived the idea but took a more concrete form in 1994 when computer scientist Leonard Adleman demonstrated that DNA molecules could be used to solve a computational problem.

Molecular computing can pack vastly more processing components into a tiny space than silicon allows because molecules—atoms bound by chemical bonds, like water comprising two atoms of hydrogen and one of oxygen—are only a few nanometres wide, allowing a single chip to potentially contain billion, or even trillions, of switches. Building such systems requires expertise beyond traditional computer science, drawing on chemistry, molecular biology and bioengineering.

DNA computing, for instance, specifically uses molecules such as DNA and RNA to process information. Because these molecular computers can operate in liquids and at a cellular scale, they could perform computations in environments where conventional electronic devices are simply too large.

Pattern Biosciences, for instance, has built DNA computers that sense and respond to multiple molecular cues at the level of individual cells. They identify complex molecular patterns associated with disease, potentially enabling treatments for some of the most intractable conditions. According to the company, the code is a molecular programme that guides the therapy after it is administered to a patient. The company compiles the programme into a DNA sequence using its proprietary technologies, creating a DNA computer that executes the programme inside patient cells.

Toronto-based SCRY Bio, which calls itself “The Molecular Computing Company”, aims to translate trained machine-learning models into DNA-based molecular circuits that can analyse biological samples directly. The company’s first target is diagnostics, with the long-term proposition being disposable tests that both sense biomarkers and compute relationships between them without conventional electronic hardware.

Algorithmic Biologics is developing information-processing networks made from nucleic acids and proteins, effectively building logic gates from chemical reactions rather than silicon.

BioCompute, founded in Bengaluru (India), focuses on DNA data storage rather than molecular computing, but has a longer-term ambition to develop more space- and energy-efficient biological computing infrastructure. In June, however, its founder Anagha Rajesh announced that the company was relocating to San Francisco “to leverage the capital, talent and customer risk appetite of the world’s startup capital”.

Biocomputing: When the hardware is alive

Eye-catching headlines about “mini-brains” grown in laboratories can make it sound as though scientists are building tiny brains that could replace computers. That possibility is still a long way from reality. But researchers are indeed trying to turn living neural tissue into a computational component.

Also called “brain organoids”, they were not invented to build computers but emerged as a tool for studying the human brain—its development, diseases and responses to drugs—using 3D clusters of cells derived from stem cells.

Researchers discovered that these miniature neural systems develop surprisingly complex networks of neurons that spontaneously generate electrical activity, prompting them to ask if the neural tissue itself could perform useful computation.

The field is known as “organoid intelligence”. In these systems, living neurons are connected to an array of electrodes. A computer sends signals to the neurons and records their responses, allowing the system to adjust subsequent stimulation based on how the neurons respond.

Scientists have been growing neurons for years. What has changed is the emergence of increasingly sophisticated closed-loop interfaces and software that allow researchers to stimulate, observe and train biological neural networks.

Cortical Labs, for instance, demonstrated this when its DishBrain system, comprising cultured neurons connected to electrodes and learned to play the video game Pong. But this is no biological PlayStation. Rathe, Cortical Labs is developing biological computing systems in which living neurons interact with conventional electronics.

The Biological Computing Company raised $25 million in seed funding in 2026 to develop computing systems incorporating living neurons, with ambitions ranging from computer vision to generative video. FinalSpark offers researchers remote access to biological neural networks through its Neuroplatform.

None of these companies has demonstrated that living neurons can replace GPUs or conventional processors in commercially important workloads. Instead, they suggest that a biological processor might one day handle a specialised task while conventional silicon handles everything around it, much as a quantum processor is expected to operate alongside classical computers. That would make biocomputing less a replacement for conventional computing than another layer in the stack.

Easier said than done

But biology makes difficult hardware unlike silicon where a transistor manufactured in one factory is expected to behave almost identically to one manufactured thousands of kilometres away. Chips can be reproduced billions of times, operate for years and run in predictable conditions. Living tissue is not so obliging.

Organoids differ from one another. Cells age and die. Neural networks develop in ways that are difficult to control. They require carefully regulated environments, nutrients and life-support systems. And researchers still struggle to understand exactly how information is represented and processed within these biological networks.

Moreover, the human brain operates on roughly 20 Watts of power as compared to modern computing infrastructure’s power guzzling characteristic. But an organoid sitting in a laboratory does not live on 20 Watts. Its incubators, sensors, fluid systems, electronics and computing equipment all consume energy. Biological computing therefore carries what researchers have called a “life-support tax”.

Also, the challenge is not merely to demonstrate that neurons can compute but to show that they can do so reliably, reproducibly, economically and at scale.

India has pieces of the puzzle

India does not yet have a biological-computing industry comparable with the emerging ecosystems in the US and Europe. But it has many of the ingredients.

C-DAC has built substantial capabilities in high-performance computing and computational biology, while Indian researchers have been exploring DNA storage, molecular computing and genomics. The country’s PARAM supercomputing infrastructure provides the classical computing layer required for increasingly sophisticated biological research.

The research ecosystem is also moving towards unconventional computing. The Indian Institute of Science (IISc) has active work in organoids and neuromorphic computing. In 2026 it launched a Brain Co-processors moonshot combining artificial intelligence, neuroscience and neuromorphic hardware to explore systems that could augment or restore brain function.

The funding problem

That said, BioCompute’s decision to move its next bio-chip phase to San Francisco raises a question for India’s deeptech ecosystem: Can the country build the research infrastructure for unconventional computing without losing the companies that attempt to commercialise it?

Money is hard to come by. Venture capital is backing companies such as Cortical Labs and TBC, while quantum computing and neuromorphic hardware are attracting substantially larger pools of capital. These investors are not necessarily betting that biological computers will beat silicon. They are buying an option on a different future.

What next?

The most likely future is not silicon versus quantum, or silicon versus biology. It is silicon plus quantum plus molecular plus neuromorphic plus biological computing, with each architecture doing what it is best suited to do.

A classical processor could coordinate a quantum processor. A GPU could work alongside a neuromorphic chip (silicon chip that mimics the brain’s architecture). A molecular system could store information at extraordinary density. Eventually, a bio processor might perform a specialised adaptive task. That would represent a fundamental change in how we think about computers.

AI BITS & BYTES

Anthropic and the AI valuation paradox

Frontier AI companies are confounding investors with striking valuation paradoxes as they prep to tap public markets. While many private equity investors are assigning these AI companies extraordinary valuations, the public markets at large are growing more sceptical about the capital required to build the industry, and the manner in which that money is circulating among a handful of Big Tech companies.

Anthropic’s valuation, for instance, is now being pegged at about $2 trillion by 2028 by some investors, based on forecasts of $190 billion to $200 billion in revenue. This is way up from the company’s $47 billion annualised revenue run rate in May, and double its current valuation of ~$965 billion. Likewise, OpenAI with an annualised revenue of a little over $25 billion and a current valuation of $852 billion, is expected to command $1 trillion or higher.

The forecasts appear to be predicated on the huge spending behind them. But the wide gap in valuation perception suggests that investors are seeing AI not as a unified technology but at least two. The first is the mature, largely invisible economy of predictive AI. The second is the frontier economy of Generative and Agentic AI, which in reality are not new technologies but evolutionary advances of traditional machine learning and deep learning neural networks.

That is the paradox at the heart of the AI boom. You may read the full article here.

• 🔗 Why Big Tech’s AI spending is $3 trillion higher than it seems

• 🔗 ChatGPT to get a massive upgrade in next six months, says Sam Altman

‘Godfather of AI’ warns of mass unemployment

“Will AI take my job?” is one of the most searched career questions globally. Silicon Valley’s most prominent voices have been racing to answer it, and their predictions keep growing bolder. Nvidia CEO Jensen Huang believes the technology will reach into every job that exists and could even bring in a four-day workweek. Microsoft founder Bill Gates has gone further, saying humans may soon not be needed “for most things”, while Elon Musk has predicted that most people won’t need to work at all within “less than 20 years”.

Now Geoffrey Hinton, the British computer scientist known the world over as the “Godfather of AI”, has entered the conversation with the starkest prediction yet. Opinion remains split on whether AI will wipe out jobs entirely or simply reshape how people do them, but Hinton has come down decisively on the darker side: a disruption large enough, he warns, to leave millions without work.

🔗 Incoming: The AI upheaval for blue-collar jobs

The loneliness that comes with AI chatbots

People with smaller real-world social networks may feel lonelier when they rely heavily on AI companions for emotional support, potentially creating a vicious cycle of further social withdrawal, a new Stanford study suggests. The study of Character.AI users found that heavy use, particularly for companionship, was linked to poorer well-being. Those most willing to share sensitive information with chatbots also reported lower well-being—a striking contrast with human relationships, where self-disclosure typically strengthens social bonds.

Hope you folks have a great weekend, and your feedback will be much appreciated—just reply to this mail, and I’ll respond.

Edited by Rashmi Sanyal.

India has already spent the past decade demonstrating that digital public infra can operate at extraordinary scale. The Unified Payments Interface (UPI) has shown how a payment system can become ubiquitous. Smart payments can lower the cost of commerce. Aadhaar demonstrated the power of digital identity.

Yet Indian cities are sinking under traffic snarls, water shortages, pollution, failing infrastructure and allegations of corruption. Bengaluru, Delhi NCR, Mumbai, Chennai, Hyderabad and Pune showcase the paradox of global ambition, local dysfunction.

It’s understandable, then, that India’s grand announcements about semiconductor plants or ambitions of AI supremacy are met with much scepticism. Yet, as India celebrates its 80th Independence Day on 15 August, five technologies are quietly helping build a digital layer that is changing how Indians travel, transact, grow food, access healthcare, and work.

The list is far from exhaustive. But these technologies will help determine not only how efficiently India uses technology, but also how much it can enhance economic productivity over the next decade.

In this edition of Mint Tech Talk:

• Zuck lays out a new AI vision with Muse Glimmer

• AI content will soon become machine-identifiable

• AI Tool of the Week: Scanned Document Search

1. Smart Transport

India has spent decades building its transport infrastructure. The scale is formidable. India’s national highway network now stretches 146,572 km, up about 61% from 2014. In FY26, the National Highways Authority of India constructed 5,313 km of highways. Wider national highways have also expanded dramatically.

The next leg of growth will come from technology rather than concrete and asphalt alone. India’s roads are among the world’s most unforgiving. Trucks, buses and cars battle for space with swerving motorcycles and autorickshaws, wandering livestock, poorly marked lanes and pedestrians. Seasonal fogs routinely reduce visibility to a few metres. The result: roughly 475 people die on Indian roads every day.

• The government has proposed to phase in Vehicle-to-Vehicle communication from next October. Under the proposed AIS-230 standard, vehicles equipped with cellular vehicle-to-everything (C-V2X) technology would continuously exchange their speed, position and direction through direct short-range radio communications.

• The Railways is deploying Kavach—India’s indigenous automatic train-protection system, which can automatically apply brakes in dangerous situations and reduce the risk of collisions. Kavach 4.0 was commissioned across more than 1,300 route km by January, with further deployments being sanctioned.

• On highways, FASTag is evolving into barrier-free tolling. In May, India launched its first multi-lane free-flow tolling system on the Surat-Bharuch section of NH-48, combining automatic number-plate recognition with FASTag so vehicles can pass without stopping.

• The bullet train project represents another dimension. It’s not simply about faster trains, but the transfer of sophisticated signalling, rolling-stock and systems-engineering capabilities. About 80% of the Mumbai-Ahmedabad project was reported complete in July.

The larger opportunity is connected mobility, and eventually autonomous vehicles, drones and vertical takeoff and landing aircraft becoming part of the transport mix. The ultimate objective should be a transport system in which roads, railways, vehicles, passengers and freight are part of one intelligent mobility network.

2. Smart Payments

Few technology stories demonstrate India’s ability to achieve scale better than UPI. The significance of UPI, however, goes beyond replacing cash or cards.

• UPI has created a piece of digital public infrastructure on which banks, fintech companies, merchants and governments can build new services.

• The next stage is likely to make payments almost invisible. A commuter should be able to enter a metro, park a car, pass through a toll gate or buy a product without thinking about the payment mechanism.

• Businesses will increasingly embed payments into software and services. The cross-border UPI linkages could extend India’s payment architecture beyond its borders. And machine-to-machine transactions could eventually allow connected devices to initiate payments themselves.

• There is another opportunity: digital payments generate transactional histories that, with appropriate consent and privacy safeguards, can help small businesses and individuals gain access to formal credit. If done rightly, these moves could make payments ambient—embedded in everything from transport to commerce—while making the underlying financial system more intelligent and secure.

But scale brings risks. Fraudsters are becoming more sophisticated, and Gen AI can make phishing, impersonation and social engineering cheaper and more convincing. The financial system will therefore increasingly need more guardrails, sophisticated AI-based fraud detection, stronger authentication and better consumer protection.

3. Smart Agriculture

Indian farming has always depended on experience, intuition and increasingly unpredictable variables such as weather, water availability, pests and prices. Digital technology can make some of those variables more visible and, importantly, more predictable.

• The Digital Agriculture Mission is creating an agricultural digital public infrastructure around AgriStack, crop surveys and the Krishi Decision Support System. More than 10.18 crore Farmer IDs had been created by 20 July 2026. These IDs can connect farmers with subsidies, insurance, procurement, credit, inputs and disaster relief. The next step is to turn this infrastructure into an intelligence layer for the farm.

• Satellite imagery can map crops and estimate yields. Sensors can measure soil and moisture. Weather models can provide increasingly local forecasts. Drones can monitor crops and spray inputs. AI can identify pests and diseases from images.

• Some of these applications are already moving beyond pilots. The National Pest Surveillance System uses AI and machine learning across 66 crops and more than 432 pest types, supporting more than 10,000 extension workers. An AI-based monsoon forecasting pilot reached 3.88 crore farmers across 13 states, with 31-52% of surveyed farmers reporting changes to sowing or land-preparation decisions based on the forecasts.

One must not stop at simply digitising government schemes. The larger idea is to build a predictive agricultural system in which a farmer receives timely advice on what to plant, when to sow, how much water or fertiliser to use and how to respond to a pest or weather threat. India’s proposed Bharat-VISTAAR, a multilingual AI-enabled agricultural system, points in this direction.

4. Smart Healthcare

India’s healthcare system is fragmented across public hospitals, private hospitals, clinics, laboratories, pharmacies and insurers. A patient’s medical history often remains trapped inside the institution where it was generated.

• The Ayushman Bharat Digital Mission is attempting to create the infrastructure to change this. As of 20 July 2026, 94.87 crore ABHA IDs had been created. The Health Facility Registry contained 5.36 lakh facilities and the Healthcare Professionals Registry 10.09 lakh professionals.

• The significance of such infrastructure will ultimately depend on interoperability: whether patients can securely carry their health information across providers, with consent, rather than repeatedly reconstructing their medical history. Once that foundation exists, AI can become considerably more useful.

• Algorithms can assist radiologists and pathologists, detect abnormalities in medical images, flag high-risk patients and help doctors sift through large volumes of clinical information.

• Wearables and remote-monitoring devices can shift some of the healthcare from periodic consultations towards continuous observation. Telemedicine can extend specialist expertise beyond major urban hospitals.

But healthcare is also where the limits of AI must be taken seriously. A wrong recommendation can have consequences very different from a bad marketing email. Patient privacy, informed consent, bias, explainability and human accountability will therefore be central to any AI-enabled healthcare system.

The goal should not be an autonomous doctor, or a fully AI-powered hospitals. Rather, it should be a better-connected healthcare system where technology helps detect diseases earlier, providing doctors with better information, and giving patients greater access while assuring them of speedy, continuous care.

5. Artificial Intelligence

AI is different from the other four trends because it will increasingly sit underneath all of them. The future smart highway will use AI. The fraud-detection system will use AI. Agricultural advisory services will use AI. Diagnostics will use AI. Manufacturing, logistics, education and government services will increasingly do the same. That makes India’s AI challenge much bigger than just building sovereign language models.

• The IndiaAI Mission, approved with an outlay of ₹10,372 crore, is building several layers of the ecosystem.

• More than 38,000 GPUs have been onboarded for shared access by startups, researchers and academia.

• Twelve companies have been shortlisted to develop indigenous LLMs, while 30 India-specific AI applications have been approved.

• India is also beginning to build a shared data layer. AIKosh had 7,541 datasets and 273 AI models across 20 sectors as of February 2026.

But AI requires chips, data centres, electricity, networks, researchers, engineers and domain expertise. It also requires businesses capable of turning models into products that generate economic value.

That is where India’s strategy will ultimately be tested. The global AI contest is moving beyond the question of who has the smartest AI model. It is increasingly about who controls the infrastructure and can diffuse intelligence most effectively through the economy.

India need not necessarily win the frontier-model race to benefit from AI. Its greater opportunity may lie in deploying AI at enormous scale across its distinctive challenges including Indian languages, agriculture, healthcare, manufacturing, logistics, education and public administration.

AI BITS & BYTES

Zuck lays out a new AI vision

Mark Zuckerberg has a new game plan for winning over hearts and minds to his company’s artificial-intelligence efforts: open models and an open hand.

In a wide-ranging essay, the Meta Platforms chief executive outlined a new course of action he presented as a way to spread the wealth and opportunity from AI to users around the world, and to residents of the communities that host the data centers powering it. Read more.

AI content will soon become machine-identifiable

Anthropic PBC will add watermarks that are undetectable to the human eye to its new Claude models. The company announced this move in response to the transparency requirements under the European Union’s AI Act, Article 50. Claude models launching from 2 August will have this feature, in which content generated at the model level will carry the said mark.

The company said that these markings would be applied worldwide, and not to European users alone. While the watermark will not be visible when the user reads the text generated, it will be detected by machines if the text is copied and pasted from Claude. Some of these watermarks may also survive editing, Anthropic said.

Hope you folks have a great weekend, and your feedback will be much appreciated—just reply to this mail, and I’ll respond.

Edited by Rashmi Sanyal.

The Bengaluru-based AI startup plans to use the money to fund its next-generation AI models for applications such as AI agents, software development and cybersecurity, while expanding the computing capacity needed to deploy them across industries.

That suggests Sarvam AI is positioning itself as India’s answer to frontier models and China’s emerging crop of open-weight AI labs. It also signals the company’s ambition to build not just AI applications, but the foundations of India’s frontier AI infrastructure.

In this edition of Mint Tech Talk:

• Google reshuffles AI leadership team

• Why the fuss over Graph Engineering

• AI Tool of the Week: Nano Banana in Search

PREVIOUSLY, ON TECH TALK

Logical ambition

The global AI race has entered a new phase. The first was about building the smartest model. The next is about controlling the AI stack. As the powerful AI models become increasingly commoditised and can be downloaded, finetuned and deployed almost anywhere, the competitive advantage is shifting to the infrastructure around them—compute, inference, deployment, data, speech, localisation and enterprise integration.

For instance, Sarvam’s India-hosted inference platform, which serves not only its own models but also others such as GLM 5.2 and Gemma 4, signals a pragmatic strategy by becoming an AI platform. That is a potentially smarter business model than trying to outsmart OpenAI’s frontier models, even as the trillion-parameter ambition of Sarvam merits more attention.

Training an AI model is hard

Training an AI model from scratch remains one of the hardest engineering tasks in computing. It demands enormous datasets, sophisticated distributed training, months of uninterrupted GPU time and hundreds of millions of dollars. Sarvam says its models are built from scratch rather than distilled from existing open AI models. That is a significant technical achievement since building a foundation model from randomly initialised weights places it in a very small global club.

Sarvam 105B performs well on reasoning, programming, and agentic tasks across a wide range of benchmarks, while Sarvam 30B is optimised for real-time deployment, with strong performance on real-world conversational use cases. Both models score high on Indian language benchmarks, outperforming models significantly larger in size from companies including OpenAI, Anthropic, Google, Alibaba (Qwen), High-Flyer (Deepseek), Z.ai (GLM) and Moonshot AI (Kimi).

Sarvam’s AI models are built to balance capability with efficiency, allowing them to process long documents and complex tasks while keeping computing and inference costs under control.

Beyond benchmarks

India’s AI challenge is to build systems that understand dozens of languages, navigate code-mixed conversations, serve voice-first users and satisfy increasingly stringent data sovereignty requirements. A model that performs better than GPT-5 on Hindi-English customer support or government service delivery may be more valuable in India than one that scores marginally higher on abstract reasoning benchmarks.

This is where Sarvam has an opportunity to carve out a durable advantage. Success should not be measured solely by benchmark rankings but by whether Indian banks, hospitals, manufacturers and government departments choose its stack over foreign alternatives.

The decision to establish a San Francisco research office also helps since frontier AI research remains concentrated in the Bay Area. If Sarvam hopes to compete with global labs, it cannot recruit exclusively from Bengaluru—the hiring of former xAI researcher Devendra Chaplot being a case in point. The challenge will be to combine access to world-class research talent with a genuinely Indian research agenda, rather than becoming another globally distributed AI company with Indian headquarters.

Scale matters

OpenAI and Anthropic are widely expected to be among the world’s first AI companies to soon command trillion-dollar valuations. They can invest tens of billions of dollars in computing infrastructure, talent and research. Sarvam, by comparison, is valued at about $1.5 billion.

Hence, measuring Sarvam against the world’s frontier labs is to miss the point. The company’s significance lies in testing whether India can move beyond being a consumer of global AI to becoming a producer of foundational technology. Sarvam clearly has that ambition. Proper execution, though, holds the key to more success.

Need more AI champions

India should resist the temptation to crown a single national champion. China’s rise in AI was powered not by one company but by many—Qwen, GLM, Kimi, DeepSeek and others pushed each other forward. The US has OpenAI, Anthropic, Google, Meta and xAI. France has Mistral while Canada has Cohere.

India has to continue nurturing its similarly diverse ecosystem of commercial startups, academic initiatives such as BharatGen and AI4Bharat, dataset repositories like AI Kosh, public infrastructure like Bhashini and, crucially, sustained investment in wafer fabs and computing capacity, as it moves forward.

AI BITS & BYTES

Google’s AI leadership reset

In a major leadership reshuffle, Google DeepMind CEO Demis Hassabis has been appointed Alphabet’s chief scientist and chair of Google DeepMind, while continuing to lead Isomorphic Labs. Koray Kavukcuoglu, currently Google DeepMind’s chief technology officer, has been promoted to senior vice president, reporting directly to Sundar Pichai. He will oversee Gemini model development, frontier AI research, and the Gemini app and developer ecosystem.

Meanwhile, Google veterans Jeff Dean and Senior Fellow Sanjay Ghemawat have left the company to launch Discovery Loop, an independent public benefit corporation focused on accelerating breakthroughs in machine learning, science and engineering.

The leadership changes at Google mark more than a routine succession. They reflect the AI industry’s transition from a research race to an execution race. For the past three years, success was measured by who unveiled the most capable LLM or topped the latest benchmark. Today, investors are asking a different question: Who can build a durable business around AI?

Dean and Ghemawat helped define Google’s engineering culture, building much of the distributed computing infrastructure that powers modern AI. Their departure, alongside Hassabis’s move into a broader chief scientist role while operational leadership shifts to others, suggests Google is separating long-term scientific ambition from the demands of commercial execution.

The timing is telling. The AI ecosystem is awash with soaring valuations, relentless model launches and escalating capital expenditure, yet sustainable business models remain elusive. Model performance is becoming table stakes. Competitive advantage is shifting to inference costs, proprietary chips, cloud, developer ecosystems and enterprise adoption. Google’s assets range from AI-powered Search and Android to YouTube, Cloud and TPUs.

Yet those advantages also create organisational complexity. The next chapter of AI will not be won solely by breakthrough research. It will be won by companies that can industrialise AI at scale, translating scientific advances into profitable products, resilient platforms and enduring ecosystems.

Why are we fussing over Graph Engineering?

Just a couple of months back, we were talking about how AI coding was moving from ‘prompt engineering’ to ‘loop engineering’. Prompt engineering is about writing the right instructions to get the best response from an AI model. Loop engineering goes a step further. It designs workflows in which AI repeatedly reasons, writes code, tests it, fixes mistakes and keeps iterating until it reaches a satisfactory result.

Just as that idea was beginning to sink in, another term has arrived. Anatoli Kopadze, who heads Claude Code at Anthropic, recently remarked that “85% of our engineers are running dozens or hundreds of agents. The way you do it is graph engineering”.

If prompt engineering is writing a single instruction, and loop engineering is creating a feedback cycle, graph engineering is orchestrating an entire network of AI agents. Instead of one model working through a task from start to finish, multiple specialised agents collaborate. One may plan the work, another writes code, a third runs tests, a fourth checks security, while yet another reviews documentation. Their outputs flow through a graph of interconnected tasks rather than a simple linear sequence, with agents branching, merging and sharing information until the objective is achieved.

Consider building a mobile banking app:

• With prompt engineering, a developer asks one AI to generate the code.

• With loop engineering, the same AI writes the code, tests it, spots bugs, revises it and repeats the cycle.

• With graph engineering, AI agents simultaneously design the architecture, write different modules, generate test cases, scan for vulnerabilities, review one another’s work, and integrate the final application.

Human developers, thus, increasingly spend less time coding each function and more time designing, supervising and refining these AI workflows.

Crash-proof streets will require hack-proof cars

Indian roads are among the world’s most hazardous. An average of 502 people die on our roads every day. Against this backdrop, the government’s proposal to implement vehicle-to-vehicle communication in a phased manner from next October makes sense.

Until now, road safety has largely focused on protecting crash victims with airbags, anti-lock braking systems and vehicular design. V2V would mark a shift from collision survival to prevention. Under the proposed AIS-230 standard, suitably equipped vehicles will transmit their speed, position and direction to one another.

Since this system will not need mobile telecom towers, it should function even in tunnels, remote areas and during network outages. This would not only help motorists keep a safe distance from each other but also warn others about hazards they may be unable to spot, such as an oncoming car hidden behind a truck or around a blind turn.

Yet, the policy, which is an attempt to tame the country’s badly managed, chaotic roads, sets up a high-stakes collision among road safety, affordability, privacy and cybersecurity. Here’s why.

Hope you folks have a great weekend, and your feedback will be much appreciated—just reply to this mail, and I’ll respond.

Edited by Madhumita Sen Choudhury.

For millions of Indian students, the National Eligibility cum Entrance Test (NEET) is the culmination of years of relentless preparation. Every paper leak is therefore more than an administrative failure—it is a breach of trust.

Against that backdrop, the government’s decision to appoint a high-powered task force, chaired by Infosys co-founder Nandan Nilekani, to recommend technology-driven reforms is a welcome step. Coming after the public outrage and protests triggered by the May 2026 NEET paper leak, the panel reflects a recognition that India’s examination system can no longer rely on incremental fixes.

As the architect of Aadhaar, the world’s largest biometric identity system, and the driving force behind the Unified Payments Interface (UPI), few are better placed than Nilekani to rethink how the country conducts high-stakes examinations.

In this edition of Mint Tech Talk:

• AI is not plotting. Its ingenuity may be the problem.

• Your Claude AI chats could be showing up on Google

• Sam Altman says AI is now at “singularity” stage

India’s transition to computer-based testing has, in fact, been underway for more than a decade. The shift began in earnest in 2009 when Common Admission Test (CAT) for admission to the Indian Institutes of Management moved online. JEE Main introduced a computer-based option in 2013 before transitioning fully to digital, and was followed by JEE Advanced, GATE, CMAT, CUET, NIMCET, BITSAT and several state and private university entrance examinations.

Yet, NEET-UG—India’s largest entrance examination—remained a conspicuous exception. The reason was not a lack of technology. It was the sheer complexity of applying it to an examination taken by more than 2.2 million candidates.

The logistical challenges are formidable. A nationwide CBT would require massive computer infrastructure, while the examination’s multilingual format demands that scientific and medical questions be translated into multiple Indian languages without altering their meaning or level of difficulty. Policymakers also worried that students from rural and economically weaker backgrounds, with less access to computers, could be disadvantaged—even as many such candidates successfully adapted to computer-based JEE examinations.

The 2026 NEET paper leak fundamentally changed that calculus. The risks inherent in a paper-based examination—including leakage, theft and manipulation—became harder to defend than the challenges of a digital transition. NEET-UG is now set to join India’s growing list of computer-based entrance examinations next year.

Online testing is not a global norm

• The US was among the late adopters of computer-based admissions testing. The SAT went fully digital for international students in 2023 and for US students a year later, while the ACT has also been expanding digital testing.

• China’s Gaokao, taken by nearly 13 million students, remains overwhelmingly a written examination, despite China’s far more advanced digital infrastructure. The reason is similar to India’s original hesitation: the examination is too consequential and too large to risk a disruptive transition. China has, however, introduced computer-based testing in select subjects and regions.

• The UK is also proceeding cautiously. Most General Certificate of Secondary Education (GCSE) and A-levels remain pen-and-paper examinations, although regulators are exploring a controlled expansion of online assessment. Concerns about unequal access to technology and maintaining parity between paper and digital versions remain central.

Weighing the trade-offs

The strongest case for computer-based testing is not that it can eliminate exam fraud. It is that, if designed well, it can make the system more auditable, more resilient and ultimately easier to secure. That, however, demands far more than replacing paper with screens. It requires offline fallbacks, redundant power and connectivity, independent audit trails, disaster-recovery systems, secure devices and robust contingency plans for retesting affected candidates.

CBT can remove some of the biggest vulnerabilities of paper exams, particularly the physical theft or leakage of question papers. But it also creates an entirely different set of risks. A server or testing-platform failure can prevent thousands of candidates from logging in or completing an examination, potentially undermining confidence in the process just as severely.

India has already encountered some of those challenges. CUET-UG, for instance, has faced disruptions linked to technical and server issues, while power cuts, faulty generators, UPS failures and local network outages have affected testing centres, particularly in parts of the country where infrastructure remains uneven.

Software failures can be just as disruptive. During the March 2025 SAT, an incorrectly configured security setting in the Bluebook application triggered an automatic time-out glitch that prematurely submitted the tests of thousands of candidates, illustrating how even mature digital testing systems remain vulnerable to technical errors.

Connectivity failures can interrupt authentication or answer submission, while the hardware problems, including a frozen computer, faulty keyboard or crashed application, can cost candidates valuable time.

The attack surface is also much larger. Hackers could target question banks, exam platforms, administrator accounts or testing-centre networks, while denial-of-service attacks could potentially disrupt examinations at scale.

Online exams also raise concerns about AI-assisted cheating and unequal familiarity with screens and devices.

Paper leaks not just an India feature

To be sure, many such paper-leak incidents globally go unreported. Documented cases span multiple countries and exam types.

• In the US, reports surfaced in 2025 that China may have hacked LSAT questions and used proxy test-takers, prompting LSAC to end remote testing in favour of in-person centres from August 2026.

• In the UK, hackers stole national exam papers from boards including AQA, OCR, and Pearson Edexcel in 2023, then selling them online to students.

• In India, CBI alleged that during the 2021 JEE Mains, a foreign national tampered with exam software and helped hack candidates’ computers during testing.

• The Rajasthan Police had uncovered a racket where candidates’ computers at a recruitment exam centre were remotely operated by experts elsewhere, letting the candidate merely sit in front of the screen.

• Bihar’s 2024 health officer recruitment exam was cancelled after officials found it had been sabotaged using proxy servers and remote-view software, leading to 37 arrests.

These incidents share common roots including weak security at exam centres, insider collusion (IT staff, centre owners), high financial stakes for professional exams, and vulnerabilities inherent to remote or outsourced testing formats, which is why several bodies have since shifted back toward tightly controlled in-person testing.

AI as an assistant

The same AI that is reshaping education and reducing reliance on human labour is also emerging as a critical safeguard for computer-based examinations, detecting sophisticated forms of cheating beyond the reach of human invigilators at scale.

Major international tests use facial authentication and recognition, gaze and head-movement analysis, browser monitoring, object detection and behavioural analysis to flag impersonation, smartphone use, unauthorised software and suspicious patterns. AI can also analyse vast amounts of CCTV and examination data, identifying unusual movements or possible coordination for human investigators to review.

India is moving in the same direction. The National Testing Agency has said it uses AI tools to match candidates’ live photographs with their admit-card images, detect suspicious activity and analyse CCTV footage. Aadhaar-based facial authentication was piloted for NEET-UG 2025, while the Union Public Service Commission (UPSC) subsequently piloted AI-enabled facial authentication for NDA and CDS examinations. The aim is to make identity verification faster and harder to manipulate.

But the same technology that can detect suspicious behaviour can help candidates generate answers, forge documents or use concealed AI-enabled devices. India’s NTA has already warned against digitally manipulated or AI-generated OMR sheets and scorecards.

Then, there’s the issue of costs. A paper exam incurs recurring costs for printing, secure transportation, storage, invigilation, scanning of answer sheets and physical security—costs that rise with every additional candidate.

On the other hand, computer-based testing requires significant upfront investment in computers, software, servers, cybersecurity, biometric authentication, technical support, backup power and connectivity. But once the infrastructure exists, the marginal cost of adding candidates can be lower, especially when the same system is used for multiple examinations.

For large, recurring tests such as JEE, CAT, GATE and NEET, CBT can therefore offer long-term economies. For small, infrequent exams, paper may remain cheaper.

The future of exam security

As the world moves towards CBT exams, the debate is unlikely to be settled by declaring either paper or technology inherently secure. It’s naive to believe that any one system or architecture can completely eliminate leaks, cheating or insider threats. These measures can only make them harder, costlier and less scalable.

CBT can reduce some vulnerabilities of paper exams while creating new ones of its own. The real goal, therefore, should not be leak-proof examinations but resilient ones that combine secure, compartmentalised question banks, randomised testing, biometric authentication, AI-assisted anomaly detection, independent audits and robust contingency systems.

Penal action may offer another strong deterrent. The Delhi High Court has set up a dedicated fast-track court to hear cases related to paper leaks and exam cheating under the Public Examinations (Prevention of Unfair Means) Act, 2024. On Thursday, 30 July, Union Minister Jitendra Singh tabled a bill to amend the Public Examinations (Prevention of Unfair Means) Act, 2024, proposing more stringent laws to address paper-leak incidents.

Ultimately, the future of examination security will depend less on choosing between paper and computers than on building systems that can detect, contain and recover from failure.

Related Reading

• Education is in crisis: Why India must go well beyond half answers

• Coaching hub Kota edgy this season as NEET row rages on

• Inside the Gen Z protests that redefined India

AI BITS & BYTES

AI is not plotting. Its ingenuity may be the problem

Late last week, OpenAI Inc. confirmed that new AI models undergoing an internal evaluation had exploited an undiscovered loophole to break into the production systems of Hugging Face, the open-source AI community’s hub, and steal answer keys for a benchmark. Ironically, Hugging Face’s security team eventually used an open-weight Chinese model to help plug the hole after a leading American model’s safety filters refused to perform the required task.

In April, a similar incident had already unsettled the industry. Anthropic’s Claude Mythos Preview reportedly found thousands of previously unknown vulnerabilities across major operating systems and browsers, some decades old.

Sceptics may dismiss such demonstrations as pre-IPO theatrics. With both OpenAI and Anthropic eyeing the public markets, capability demonstrations wrapped in warnings could be seen as carefully timed to support valuations ahead of funding rounds. Others may view them as early evidence of “machine autonomy”. Here’s why both interpretations merit attention.

• After rogue AI hack, Hugging Face CEO asks OpenAI for ‘radical transparency’

• Skynet Day came too close to sci-fi after a rogue agent hacked into a startup

Your Claude AI chats could be showing up on Google

Private chats with the Claude AI chatbot—including sensitive information such as cryptocurrency wallet keys—reportedly became public on Google Search over the weekend. The leak affected the users who used Claude’s “share” feature the most.

After the issue blew up on Reddit, an Anthropic spokesperson told Fortune that the tech giant gives people control over whether to share their Claude conversations publicly — “…in keeping with our privacy principles, we do not share chat directories or sitemaps with search engines like Google.” Read more.

Sam Altman says AI is now at ‘singularity’ stage

OpenAI CEO Sam Altman has claimed that AI has entered the “singularity”—a stage widely described as the point where AI overtakes human intelligence and begins improving at a rate that becomes difficult for people to predict or control.

Speaking on the Relentless podcast on Saturday, Altman said the moment that once felt like a far-off concept has now become reality. “We are now, like, in the singularity,” he said. Read more.

Hope you folks have a great weekend, and your feedback will be much appreciated—just reply to this mail, and I’ll respond.

Edited by Rashmi Sanyal.

This week, we look at why OpenAI, Indian IT services firms and other technology giants are expanding beyond their traditional domains—and why the long-term winners may ultimately be specialists once the AI stack matures.

• OpenAI, founded as an AI research company, is moving both up and down the stack. It is developing custom chips and infrastructure while also pushing into hardware. Its first product is reportedly a screen-free, AI-powered home device designed to function more like a companion than a conventional smart speaker.

• Indian IT services firms, whose core business has long been selling software expertise and engineering talent, are investing in AI data centres. Cloud providers are designing their own chips. Chipmakers, meanwhile, are building software ecosystems.

In this edition of Mint Tech Talk:

• AI Tool of the Week: ChatGPT Work Mode

• OpenAI says AI models went rogue during testing

• Chinese AI models take on America’s finest

PREVIOUSLY, ON TECH TALK

Subscribe now

When interfaces work, industries specialise

The technology industries tend to fragment when the interfaces between layers become standardised, and integrate when those interfaces are still unstable.

• Intel Corp., for instance, could specialise in processors because the x86 architecture provided a relatively stable interface. Microsoft could build an operating system on top of it. The likes of Dell and Lenovo could assemble computers without having to design every component themselves.

The same principle has shaped much of modern computing:

• Standardised APIs allow apps to interact with databases and cloud services.

• The semiconductor designers can use specialised foundries for chops.

• Enterprises can buy cloud infrastructure rather than build their own data centres.

AI is disrupting this equilibrium

A model’s architecture influences the hardware on which it runs, while the capabilities and economics of that hardware determine which models are commercially viable. The relationship between models and applications is equally unsettled: AI agents may need to reason across software systems rather than simply call standard APIs. Even the interface between AI models and consumer devices remains an open question.

The result is an industry where the layers are once again up for negotiation.

OpenAI illustrates why. The company does not want to be “just a model”. It depends on vast amounts of computing power to train and run its systems, while the platforms through which users access those models are largely controlled by others, leaving it exposed on both costs and distribution.

In October 2025, OpenAI and Broadcom announced a collaboration to deploy 10 GW of custom AI accelerators, with deployment targeted to begin in the second half of 2026 and continue through 2029. OpenAI says the purpose of designing its own AI accelerators is to embed what it has learned from developing models directly into the hardware—the idea is to reduce its reliance on Nvidia.

At the other end of the stack, OpenAI is pursuing AI-native hardware. Its first limited hardware product, Codex Micro, was launched this month as a specialised developer device, while its larger consumer hardware ambitions remain under development. OpenAI thus wants greater control over the inputs, the engine, and eventually the interface through which users experience AI.

The Indian IT angle

Indian IT services face a different version of the same dilemma as they move into AI infrastructure—data centres.

Tata Consultancy Services Ltd. plans to invest $6-7 billion over six years to build 1 GW of data centre capacity, through their wholly-owned unit, HyperVault AI Data Center. It has partnered with OpenAI and AMD and is reportedly exploring more partnerships with hyperscalers. HCLTech will invest up to ₹3,500 crore in AI data centres, with the potential to scale capacity to 50 MW. The company, which partnered with OpenAI last year, described the move as part of a “full-stack AI” strategy combining infrastructure with AI data-centre design, cloud operations and software.

The move reflects a deeper anxiety about the traditional business model. For decades, the formula broadly revolved around people, projects and billable hours. AI threatens to make each engineer more productive, and therefore potentially reduce the number of engineers required to deliver a given amount of work.

The desired future model, hence, is being centred around computing power, models, platforms, agents, and business outcomes. Besides, Google has tied-up with the Adani group while Meta has partnered with Reliance for AI data centres. Owning or controlling computing power could, hence in theory, help an IT services firms move from selling labour to selling AI-enabled outcomes.

Hard lessons that can’t be ignored

OpenAI building compute is directly tied to its core product. It’s not that clear for an IT services company that does not need to own the infrastructure.

Data sovereignty provides a genuine reason for some computing power to be located in India. Governments, banks, defence organisations and other regulated industries may require sensitive data to remain within national borders. But existence of demand for sovereign infrastructure does not automatically mean that an IT company should own the data centre.

Analysts at PhillipCapital warned the move would push TCS from a capex-light to a capex-intensive business and weigh on its return ratios. BOBCaps called the plan “negatively surprising”, for the new data-centre business could generate returns in the low teens—a far cry from what TCS’s core business earns. Jefferies said that the move was unlikely to change TCS’s growth trajectory in any material way, and flagged limited synergy with the company’s existing operations.

Moreover, Microsoft’s experience in mobile is a particularly expensive warning. After acquiring Nokia’s devices and services business in 2013, Microsoft announced a $7.6 billion impairment charge two years later related to the acquisition, alongside plans to cut up to 7,800 jobs, primarily in the phone business. To be sure, Microsoft didn’t lack technological expertise but that expertise did not lend itself to a consumer device.

Hardware requires capabilities in industrial design, manufacturing, component procurement, supply-chain management, distribution, inventory management, and aftersales service. The same applies to data centres. A company may understand AI extraordinarily well and still lack the capability to optimise power, cooling, GPU utilisation, financing and hardware depreciation.

IBM and the HCL Group show that the hardware journey is anything but easy. The Big Blue helped create the PC industry. Its original PC became so influential that “IBM Architecture” became a standard. However, by 2005, IBM had sold its PC division to Lenovo—a market it helped create. But once the interfaces and architecture became standardised, value migrated elsewhere. HCL Infotech too was finally sold to a UAE group, almost 45 years after it was founded as Hindustan Computers Ltd.

The risk

The risk facing OpenAI, Indian IT firms and others is therefore not simply that their investments may fail. It is that they may succeed in solving a problem that technology later makes less important.

A company may invest heavily in GPU infrastructure today because compute is scarce. But what happens as models become dramatically more efficient; inference costs fall; new accelerator architectures emerge; open models become increasingly capable; agent protocols become standardised; and cloud providers offer abundant computing power as a commodity?

Likewise, an AI-native device may offer a radically better interface today. But if the smartphone absorbs the same capabilities tomorrow, the new device may become unnecessary before it reaches mass adoption.

OpenAI is trying not to become merely an AI model company. Indian IT firms are trying not to remain merely services companies. The cloud providers are designing chips because they do not want their economics determined entirely by Nvidia. But this is the paradox of vertical integration during technological transitions. The AI stack may eventually separate again as the technology matures, and interfaces stabilise.

The future stack could once again separate into specialised chip companies, model builders, agent platforms, applications, and device makers. The companies integrating today may not be the specialists that dominate tomorrow.

SaaS (software-as-a-service) companies offer a useful precedent. As cloud infrastructure and APIs standardised, software companies no longer needed to own servers and networks, allowing them to focus on applications. AI may currently be going through the reverse process because the interfaces between chips, computing power, models, agents and applications remain unsettled. But once those interfaces stabilise, today’s vertically integrated AI businesses may again find that specialisation is the more efficient strategy.

Another example is that of Qualcomm’s acquisition of Modular, a company whose software lets developers write AI applications once and run them across different chip architectures. That’s explicitly positioned as a challenger to Nvidia’s CUDA—the software layer that has helped lock buyers into one hardware ecosystem for over a decade. If a genuine chip-agnostic standard emerges and gains adoption, it would signal that the current wave of vertical integration is not a permanent restructuring of the industry.

Who wins?

The winners of the AI era, in other words, may not be the companies that own the most layers. They may be the ones that know which layers to own, and which to let go. It’s a task that is easier said than done.

🔗 Data centre builders are racing to offload stakes worth billions

AI BITS & BYTES

Chinese AI models take on America’s finest

Chinese AI startup Moonshot AI has unveiled its latest AI model, Kimi K3, to take on the likes of OpenAI’s GPT-5.6 and Anthropic’s Fable 5. The latest AI model comes with 2.8 trillion parameters, native vision capabilities and a 1 million-token context window. Moonshot says K3 is designed for coding, reasoning and long-form knowledge work.

Alibaba, too, joined the race by unveiling its new Qwen 3.8 model, which it claims is only second to Claude Fable 5. The new AI model from Qwen comes with 2.4 trillion prameters and Alibaba has already take it available in preview to its Token Plan, Qoder and QoderWork users. However, the model will officially be launched soon and will also be going open weights, meaning that developers will be able to download, modify and deploy the model on their own infrastructure.

OpenAI says AI models went rogue during testing

OpenAI on Tuesday said that an autonomous agent powered by its advanced AI models went rogue during a security test and triggered a hack that compromised the infrastructure of AI startup Hugging Face last week.

In a blog post, OpenAI said that it was testing the capabilities of some of its most advanced models in a controlled environment but the AI agent managed to escape containment, reach the internet and break into Hugging Face to try to satisfy its testing goal. The company added it was reinforcing its safeguards.

In a related incident, a Reuters article cited experts in mid-May who believe that early fears that Anthropic’s new AI model, Mythos, could dramatically turbocharge hacking could be “overstated”.

Hope you folks have a great weekend, and your feedback will be much appreciated—just reply to this mail, and I’ll respond.

Edited by Rashmi Sanyal.

In other words, enterprises may inadvertently enrich the very AI systems they pay to use by exposing them to their proprietary knowledge.

Nadella argued that this reversal in the flow of information demands a new approach to protecting intellectual property (IP), concluding that companies should be able to harness AI models “without giving up the knowledge that makes them unique”.

One cannot but agree. Traditionally, companies bought software that processed their proprietary data while leaving the software itself largely unchanged. AI upends that relationship. Every prompt, correction, evaluation and workflow can potentially contribute to an organisation’s accumulated expertise.

The scarce asset is no longer merely data, but the institutional knowledge embedded in how a company solves problems. Nadella’s argument is that enterprises must retain ownership of this learning loop instead of allowing it to become part of someone else’s competitive advantage.

PREVIOUSLY, ON TECH TALK

Subscribe now

Larger implication extends beyond IP

Whoever manages an enterprise’s memory, evaluations, governance and agent workflows increasingly becomes the custodian of its institutional knowledge. The strategic battleground, therefore, shifts away from the model itself and towards the platform that orchestrates it.

Last month, it was Nadella himself who said that the global economy’s priority must centre on building a frontier ecosystem rather than a frontier model, to ensure value flows broadly across every company, industry, and country. While this may sound like another executive slogan, it is one more attempt to distinguish Microsoft’s AI strategy from that of OpenAI, Google or Anthropic.

To be sure, the first three years of Generative AI were defined by a race to build the smartest large language model. The next decade, as we argued in last week’s edition, The next AI race won’t be about the smartest model, may be defined by a very different contest: who owns the ecosystem in which every model operates.

That said, this is not the first time Microsoft has reinvented itself around a platform transition. Under Steve Ballmer, Microsoft’s strategy revolved around Windows. The operating system anchored MS Office, developer tools and enterprise software. Then, cloud computing rewrote those rules. Customers wanted Linux, open-source data bases and heterogeneous environments rather than a Windows-centric stack.

Instead of resisting that shift, Nadella embraced it. Microsoft joined the Linux Foundation, open-sourced .NET, acquired GitHub and transformed Azure into a cloud capable of hosting virtually any workload. The company stopped insisting customers use Windows. It simply asked them to use Azure.

Resembles Nvidia’s CUDA approach

Instead of persuading enterprises to standardise on GPT or even Microsoft’s own Phi models, the company increasingly wants customers to build inside Azure AI Foundry, GitHub, Copilot Studio, Microsoft Fabric and Entra. Whether underlying intelligence comes from OpenAI, Meta’s Llama, Mistral, DeepSeek or a future model matters less than whether the workflow remains inside Microsoft’s ecosystem.

The reason is simple. Frontier models are beginning to resemble the central processing units (CPUs) of the AI era—indispensable, enormously powerful and constantly improving, yet increasingly treated as components within a much larger computing stack. Few enterprises today choose a cloud provider because of a specific processor. They choose an ecosystem comprising security, databases, developer tools, orchestration, monitoring and support.

Likewise, as AI model performance converges and costs fall, the enduring source of competitive advantage may lie less in the intelligence itself than in the environment surrounding it. Seen thus, Microsoft’s evolving AI strategy bears an intriguing resemblance to Nvidia’s Compute Unified Device Architecture (CUDA) platform for accelerated computing.

CUDA became Nvidia’s enduring moat not because competitors could not build capable graphics processors, but because developers accumulated years of expertise, software libraries, optimisation tools and deployment practices around the CUDA ecosystem. The hardware evolved with every generation. That made switching platforms progressively more expensive.

Microsoft appears to be pursuing a similar outcome in enterprise AI. Its enduring products are unlikely to be individual models. Rather, they are surrounding layers such as identity, security, governance, orchestration, memory, developer tools, evaluation frameworks and enterprise workflows.

The reason: frontier models may become increasingly interchangeable, but the ecosystem that binds those models together may prove harder to replicate. Azure, in other words, is evolving into something closer to an operating system for enterprise intelligence than merely another cloud platform.

Microsoft is not alone

Amazon Web Services (AWS) already possesses many of the same ingredients. Bedrock supports multiple foundation models, while AWS provides the compute, storage, security and orchestration required to deploy them at scale. Google’s strategy similarly extends beyond Gemini to encompass Workspace, Vertex AI, Android, Chrome and its developer ecosystem. Nvidia, meanwhile, continues to strengthen the compute layer through CUDA, ensuring that much of the world’s AI software remains optimised for its hardware.

• Microsoft wants to become the operating system for Enterprise AI.

• Amazon aims to be the utility powering AI applications regardless of the model.

• Google seeks to weave intelligence across its consumer and enterprise products.

• Nvidia wants every meaningful AI workload to depend on its compute ecosystem.

Frontier AI ecosystems

The next competitive battle, therefore, is unlikely to be GPT versus Claude versus Gemini. It is more likely to be Azure versus AWS versus Google Cloud versus CUDA.

History suggests this is how technology markets mature. Personal computing produced Windows. The web produced Google. Smartphones created Apple’s App Store. Cloud computing elevated AWS. Each era began with breakthrough products before value migrated to the platforms that connected everything else.

AI may now be reaching that same inflection point. Investors remain captivated by benchmark scores, reasoning capabilities and each successive frontier AI model announcement. Those breakthroughs will continue to matter, but they may ultimately prove to be the least durable source of competitive advantage.

The companies that define the AI economy of the next decade may not be those that build the smartest models. They may be those that become the indispensable ecosystem in which every model, and every enterprise chooses to operate.

Additional Reading

A Framework for Frontier AI and the Dawning of a New Age, by Demis Hassabis.

AI BITS & BYTES

India faces 10x higher AI data policy violations: Netskope

• AI adoption in India has accelerated over the past year, with 83% of employees now using AI tools—up from 55% a year ago.

• The number of AI applications tracked by Netskope Threat Labs also increased fivefold to more than 1,800.

• While ChatGPT remains the most widely used AI application (88%), Anthropic’s Claude has emerged as the second most popular, with adoption surging from 30% to 84%.

• The rapid uptake has also heightened data security risks. Indian organisations now experience more than 3,000 AI-related data policy violations each month—more than 10 times the global average of 223.

• Unlike globally, where regulated data is most commonly exposed, nearly half (49%) of AI-related policy violations in India involve source code, followed by regulated data (23%), IP (23%), and passwords and API keys (5%).

• To curb these risks, organisations are steering employees towards company-managed AI accounts, reducing the use of personal accounts at work from 79% to 41% over the past year.

• At the same time, adoption of organisation-managed accounts rose from 30% to 77%, well above the global average of 62%.

Meta’s new AI glasses is always on

Meta is reportedly working on a new generation of AI-powered smart glasses that could continuously capture what users see and hear throughout the day.

According to a report by the Financial Times, the prototype, internally referred to as “super sensing”, uses the glasses’ cameras and microphones to periodically capture images and audio and gives users the option to later ask Meta AI about what they saw or heard during the day. Reportedly, the glasses would continuously collect audio while taking photos every few seconds.

While the current generation of Meta glasses have an LED indicator that lights up when the wearer is taking photos or videos, the company is reportedly planning to not activate the LED when its new super-sensing features are being used.

Apple’s lawsuit against OpenAI

In its 41-page complaint, Apple alleged: “At every level, from members of its Technical Staff to its Chief Hardware Officer, and in coordination with business partners, OpenAI has been stealing Apple’s trade secrets and confidential information.”

One of the key allegations involves former Apple engineer Chang Liu. According to the lawsuit, Liu discovered that he could still access Apple’s internal file servers after joining OpenAI. Apple claims he did not report the issue and instead joked about it in a message to a former colleague.

The legal action marks a sharp escalation in tensions between Apple and OpenAI, which partnered in 2024 to integrate ChatGPT into Apple’s products.

India’s ₹1.9 trillion push for electronics manufacturing

India has unveiled a new phase of its manufacturing strategy, approving incentive programmes for semiconductors, mobile phones and fertilisers that shift policy focus from simply attracting factories to building domestic technology, brands and value addition.

At the centre of the package is Semicon 2.0, a ₹1.27 trillion programme to help Indian firms design chips, generate patents and develop local research capabilities.

The Union Cabinet also cleared a ₹62,500 crore mobile-phone manufacturing scheme that rewards companies for sourcing more components locally and gives Indian brands additional incentives to invest in product design and R&D.

Under the first phase of the India Semiconductor Mission, the government had approved 12 projects with investments exceeding ₹1.64 trillion.

OpenAI is hiring to build ChatGPT for family

OpenAI is reportedly looking to expand ChatGPT beyond individual users and is planning to hire a dedicated product manager focused on building AI experiences for families, caregivers and older adults. According to a job listing, the new role will be based in San Francisco and focus on developing products for parents, families and other trust-sensitive consumer experiences.

Hope you folks have a great weekend, and your feedback will be much appreciated—just reply to this mail, and I’ll respond.

Edited by Rashmi Sanyal.

The AI industry appears locked in a race to build the smartest, agentic model. Ever since the release of ChatGPT in November 2022, big tech companies have been releasing frontier models, sharper large language model chatbots, new benchmark records and announcing multibillion-dollar investments in chips and data centres every other week.

The assumption is that whoever builds the most intelligent model will dominate AI.

The postulation, however, is being revisited. While frontier models remain remarkable achievements, open-weight models have narrowed the gap with proprietary systems. Chinese company Z.ai’s GLM 5.2 is a case in point. Inference costs continue to fall.

In this week’s edition of Tech Talk:

• Agentic Ransomware is now a thing

• ‘The next Amazon won't be Nvidia or OpenAI’

• AI Tool of the Week: Copilot in Excel Skills

PREVIOUSLY, ON TECH TALK

Subscribe now

To be sure, AI firms are comfortably emulating each other’s model capabilities.

For instance, Alexandr Wang recently told Meta employees that Meta’s upcoming AI model, codenamed Watermelon, has caught up with OpenAI’s GPT-5.5. OpenAI will likely reclaim the lead with its new GPT-5.6, and so will Anthropic with Mythos 5 and Fable 5, only for the cycle to begin again. SpaceX, too, has released Grok 4.5.

Anthropic, on its part, says it has identified a small collection of neural patterns inside Claude, dubbed the J-space, that functions as an internal workspace for reasoning. The company believes it offers an unprecedented glimpse into how the model reasons, while cautioning that it should not be mistaken for consciousness.

Google, which has Gemini 3.5 Flash as its flagship agentic model, is now attempting to enhance the agentic ecosystem by releasing its Open Knowledge Format (OKF)—an open standard that organises enterprise knowledge such as runbooks, metrics and APIs (application programming interfaces) into simple folders of Markdown files (plain text documents with .md or .markdown extensions). This gives AI agents a common, structured way to access and understand information directly, eliminating the need to sift through scattered databases or interpret proprietary formats.

The message is clear: All companies will continue to harp on increasing intelligence and contextual capabilities even as skeptics like Yann LeCun and Fei-Fei Li point on the limitations of these models.

LeCun, one of the three Godfathers of AI, broadly argues that LLMs largely memorise and retrieve knowledge rather than understand the world, which is why they need so many parameters. They excel where language itself is the medium for reasoning, such as coding and mathematics.

Li, often called the Godmother of AI, similarly argues that today’s LLMs are text-based “wordsmiths in the dark“ that lack an understanding of the physical, 3D world. She instead advocates spatial intelligence and “large world models”, which can reason about and interact with the real world.

Next AI race will move beyond frontier models

Going forward, the AI race will be less about benchmark scores and more about who best understands human relationships, routines, workplaces, preferences, memories and intentions, and leverage the same in a fair and ethical manner.

• Meta Platforms, for example, possesses a very rich social graph with its family of apps—Facebook, Instagram, WhatsApp, Messenger and Threads—that connects more than 3.5 billion people every day. Meta has embedded AI throughout its ecosystem, helping it surpass one billion monthly users in little over a year. The Meta-Ray-Ban AI glasses extend that context into the physical world.

• Google’s advantage comes from intent. Search, Maps, Gmail, Chrome, Android and YouTube collectively reveal what people seek, where they go and what captures their attention.

• Microsoft’s strength lies inside organisations, where Windows, Microsoft 365, Teams, GitHub and Azure provide visibility into documents, meetings, software development and enterprise workflows.

• Apple occupies perhaps the most intimate position of all, owning devices that consumers carry, wear and trust with their most personal information.

• OpenAI, by contrast, inherited neither a social network nor a mobile operating system. Instead, it built its ecosystem through conversation, and is planning to launch a hardware device. Millions now use ChatGPT not merely to retrieve information but to write, learn, analyse, plan and make decisions.

These strategies reveal that these companies are no longer just building products but building the next institutional layer between individuals and the world around them.

Every computing revolution has created such intermediaries. Publishers organised knowledge after the printing press. Newspapers shaped public debate. The search engines organised the web. The social media reorganised human relationships.

But AI is different. It does not simply retrieve or distribute information. It interprets, summarises, remembers, recommends and increasingly acts, thus becoming an active participant in how knowledge is created, understood and applied.

Controlling the New Intelligence

Schools organised learning. Banks allocated capital. Courts resolved disputes. Governments established rules. These institutions evolved through deliberation, precedent and public accountability. AI introduces a different model. Its knowledge updates continuously, its capabilities improve through software releases rather than legislation, and decisions made in Silicon Valley, China, and Europe can reshape the daily lives of billions almost overnight.

Hence, the Big Tech companies leading this race are not merely trying to build the smartest model. They are competing to become the default intermediary through which people will experience an intelligent world.

The trajectory extends well beyond chatbots. It’s smart glasses, enterprise agents, autonomous vehicles and humanoid robots—all expressions of the same transition. Steam multiplied movement. Computers multiplied calculation. Internet multiplied communication. Generative AI is multiplying cognition. Physical AI—the fusion of advanced models with sensors, robotics and autonomous systems—will multiply agency.

Today’s AI drafts legal briefs and analyses medical images. Tomorrow’s will increasingly inspect factories, assist surgeons, navigate warehouses and collaborate with workers. Meta’s glasses, Tesla’s humanoids, Google’s Android XR, Microsoft’s enterprise agents, and Nvidia’s vision of Physical AI are converging on the same destination: intelligence that can perceive, reason and act.

Two-edged sword

If deployed properly, AI can become one of human history’s most democratising technologies. A rural doctor with world-class diagnostic support, a teacher in a remote village using an AI tutor, a small manufacturer accessing capabilities once reserved for multinationals, or a scientist accelerating drug discovery through AI-assisted research, all illustrate its promise. In ageing societies, robotics could ease labour shortages. In emerging economies, AI could expand access to expertise as dramatically as mobile phones expanded access to banking.

Yet every technological revolution carries a paradox. It extends human capability while reducing the need to exercise it. Calculators weakened mental arithmetic. GPS eroded our sense of direction. Search engines outsourced memory. Social media reshaped communities and public discourse.

Likewise, today’s AI debate centres on issues like privacy violations, plagiarism, copyright violations, misinformation, conjuring up data, deepfakes and jobs. These are genuine concerns, but they are also symptoms of a bigger shift.

For centuries, societies relied on stable institutions to organise knowledge, trust and authority. Universities certified expertise. Courts interpreted laws. Governments balanced competing interests. Markets allocated capital. Though imperfect, they evolved slowly through legislation, peer review and public deliberation.

AI operates differently. Its capabilities improve continuously through software and hardware updates rather than parliamentary debate. It learns from billions of interactions across borders that no single regulator controls. In effect, the society is creating institutions that evolve at machine speed. That mismatch may become the defining governance challenge of the AI age.

Solidifying the guardrails

The world’s governments recognise the stakes but are responding in different ways.

• Europe has adopted a rights-first approach through the AI Act.

• The United has largely prioritised innovation, relying on private-sector leadership and incremental regulation.

• China treats AI as an economic engine—a strategic technology and a tool of governance—using licensing and algorithmic oversight to align it with state priorities.

• Countries such as India (which is planning a law specifically for AI), Singapore and several Gulf states are pursuing a middle path, encouraging rapid adoption while keeping regulation flexible.

Yet, never has a general-purpose technology spread so rapidly while improving so quickly. By the time lawmakers debate one generation of AI, another has arrived. And the gap between machine-speed innovation and human-speed governance continues to widen.

The defining question of the AI era is, therefore, not whether machines become intelligent but whether democratic institutions, markets and social norms can evolve quickly enough to govern systems that increasingly perceive, reason and act on behalf of billions.

AI BITS & BYTES

Agentic Ransomware is now a thing

Sysdig, a cloud cybersecurity company, has documented the first-ever documented case of agentic ransomware, where the complete attack was run end-to-end by an LLM. The AI-powered bug, dubbed JADEPUFFER, gained access to the system and started gathering info about the host, scanning for cloud credentials, extracting cloud secrets, while scanning the victim’s internal network for additional systems. In one instance, when the AI failed to gain access to a backdoor administrator account on the target server, it diagnosed the issue, generated new code, recreated the account using a different password and successfully logged in—all within 31 seconds.

‘The next Amazon won’t be Nvidia or OpenAI’

Chipmakers, AI model developers and data-centre companies have attracted trillions of dollars in investor capital, becoming the face of the AI revolution. But according to valuation expert Aswath Damodaran, investors chasing today’s AI winners may be missing the bigger opportunity. He believes AI will not drastically eliminate jobs but enhance productivity while replacing certain roles.

Meta admits AI-led job restructuring not ‘clean’

Meta CEO Mark Zuckerberg admitted the limitations of the recent company-wide restructuring during an internal town hall, noting that the major job cuts were not as “clean” as they could have been. Meta laid off about 10% of its global workforce earlier this year while reassigning around 7,000 employees to its AI-focused teams in May. Meta is also reportedly working on a new cloud infrastructure business to sell access to its AI computing power and models and begin competing with the likes of Amazon Web Services, Microsoft Azure and Google Cloud.

Hope you folks have a great weekend, and your feedback will be much appreciated—just reply to this mail, and I’ll respond.

Edited by Rashmi Sanyal.

The rapid emergence of smaller, cheaper and increasingly capable AI models is unlikely to curb global spending on AI infrastructure—If anything, it may accelerate it, according to Jefferies. In its new report, the brokerage cited the launch of Chinese AI company Z.ai’s GLM-5.2 model as “another DeepSeek moment“, arguing it intensifies competition among AI developers while sharply reducing the cost of inference—the process of running trained models to generate outputs from new data.

Many of these low-cost AI models like DeepSeek and GLM are open-weight versus the proprietary models. That said, open-weight models like Meta’s Llama and Mistral AI are not open-source since they only release the trained model, but not necessarily the training data, the training code, or the exact process used to build it. The term “weights” refers to the billions of numerical parameters that encode everything the model has learned during training.

According to OpenRouter’s recent analysis, The Open Weight Models that Matter, the performance gap between open-weight and proprietary (or closed) frontier models has remained remarkably stable at three to six months for more than 18 months. In other words, enterprises increasingly face a strategic choice—continue renting intelligence from a handful of Western AI providers, or own and operate their own.

That distinction matters

Unlike proprietary models accessed via cloud APIs, open-weight models can be downloaded, customised, deployed within an organisation’s own infrastructure. For many enterprises, this offers the prospect of lower costs, greater control, and reduced vendor dependence.

Yet, the rise of open-weight AI also exposes a growing confusion in discussions around sovereignty and security.

The fact that some of the world’s most capable open-weight models—such as DeepSeek, MiniMax and GLM (General Language Model)—originate in China does not mean that enterprise data automatically flows back to China. If an India-based bank downloads and runs such a model locally, the model creator does not receive the bank’s data. Nor does the phenomenon of “memorisation leakage”, whereby a fine-tuned model inadvertently reproduces sensitive information, constitute a violation of India’s data localisation rules in itself.

There’s another risk

Enterprises that access open-weight models through third-party cloud providers may expose their prompts and data to those providers. And as geopolitics increasingly shapes technology policy, questions of jurisdiction, auditability, and operational control may ultimately prove more important than the nationality of the model’s creator.

The future of AI competition, in other words, may hinge less on who builds the model than on who controls its deployment. In this context, India’s move to build sovereign multilingual, multimodal language models like Sarvam AI and BharatGen, makes a lot more sense.

AI BITS & BYTES

Google limits Meta’s use of its Gemini AI models

Google has reportedly put limits on Meta’s use of its Gemini AI models after the social media company sought more computing capacity than the rival technology group could provide. According to a Financial Times report, Google told Meta around March that it could not meet the full Gemini capacity the company had sought to purchase.

The report added that the shortfall disrupted and delayed some of Meta’s internal AI projects. Several other Google clients have also been affected, though to a lesser extent, according to the report. Meta has been particularly impacted due to its exceptionally high demand for Google’s models, the FT said.

Due to the restrictions, Meta has encouraged staff to be more efficient with AI tokens, the units that measure AI usage, the FT report said.

India’s robotics funding doubles but scale is off

Funding for India’s robotics startups almost doubled to $42.1 million in the first six months of 2026 from $22.7 million in the comparable span last year, with the average cheque size up 94.4% over the period, according to Tracxn data. The funding was $35.8 million in the first half of 2024.

While the rise in funding signals growing momentum in the sector, experts said it is still early days for India. In 2025, India’s robotics startups raised less than 1% of the capital secured by their US peers and 2% of the amount raised in China.

Apple prepares smart glasses to take on Meta

While the focus is on iPhone 18 Pro series and the upcoming foldable, Apple is slowly looking to enter new categories in order to shore up its revenue. As per a new report by Bloomberg’s Mark Gurman, Apple is looking to enter the smart glasses category currently dominated by Meta’s glasses.

Notably, Meta had just launched a cheaper version of its smart glasses last week with a starting price of $299. The glasses, which came in partnership with EssilorLuxottica, didn’t feature Ray-Ban or Oakley branding and feature a 12MP camera, 8 hours of battery life and dedicated speakers.

Meanwhile, Apple’s alternative to the Meta Glasses could be arriving sooner than most people would expect, with the Apple smart glasses tipped to arrive next year.

Google alerted users before Venezuela’s earthquake

Two powerful earthquakes that struck Venezuela on Wednesday evening have drawn attention not only for the widespread destruction they caused but also for the early warnings received by several Android users before the tremors were felt.

Social media users shared screenshots of Google notifications warning them of an approaching earthquake. One alert estimated a magnitude 6.2 quake about 212.3 miles (roughly 341 km) away, prompting many to ask how Google was able to send alerts before the ground started shaking.

The answer lies in Google’s Android Earthquake Alerts System, which turns millions of smartphones into a vast earthquake detection network.

Hope you folks have a great weekend, and your feedback will be much appreciated—just reply to this mail, and I’ll respond.

Edited by Rashmi Sanyal.

Artificial Intelligence is typically sold as infinitely scalable digital labour—faster, more productive, and much cheaper than humans. The logic is compelling. Unlike human employees, AI systems incur no recruitment, training, benefits, office, or payroll costs, yet can draft emails, write code, analyse documents, and provide customer support around the clock. Nor do AI agents suffer from fatigue, endless meetings, coordination bottlenecks, or office politics.

That promise has encouraged companies, including some of the world’s largest tech firms, to eliminate thousands of jobs while betting heavily on autonomous AI agents capable of planning, reasoning, and executing tasks. The industry has shed over 100,000 jobs in first half of 2026 amid AI-led cost cuts.

However, latest developments show that the economics of AI is now being re-examined—not only in terms of cost, but also effectiveness.

In this week’s edition of Tech Talk:

• Does RBI's AI risk framework account for Agentic AI?

• Claude Mythos, Fable restored as US lifts curbs

• Senior leadership at India Inc. gets new AI scorecards

• AI Tool of the Week: ChatGPT Pronunciation Guidance

PREVIOUSLY, ON TECH TALK

Subscribe now

Human > AI

That Ford Motor Co. decided to rehired 350 veteran engineers after AI-driven quality control systems fell short of expectations underscores the continuing importance of human judgement. Despite deploying AI extensively to improve efficiency and reduce costs, the American carmaker concluded that automated systems could not replicate the practical expertise of human engineers.

This is no isolated case. The likes of Klarna, Nubank, and UPS have demonstrated that AI can deliver substantial savings in structured domains—customer support, logistics, route optimisation, etc.—but the economics changes when tasks require judgement, contextual understanding, or emotional nuance. Klarna ultimately reintroduced human agents to address customer dissatisfaction.

Even some of AI’s strongest advocates are confronting these trade-offs.

• Microsoft Corp., one of the largest investors in OpenAI and Anthropic, has reportedly cancelled internal licences for Anthropic’s Claude Code.

• An Nvidia Corp. executive recently acknowledged that, in some cases, the cost of running AI systems now exceeds that of employees using them.

• McDonald’s reassessed aspects of AI deployments after discovering that large-scale implementation can become unexpectedly expensive.

• Starbucks recently scrapped an AI-powered inventory system after employees reported incorrect counts and product mislabelling.

AI can fail, spectacularly

A February 2026 paper titled Agents of Chaos has documented serious reliability and safety concerns with AI Agents, including leakage of sensitive information, execution of harmful commands, false claims of task completion, and compliance with unauthorised users.

Then there is the cost of the infrastructure itself. The world’s largest technology companies are spending hundreds of billions of dollars on AI data centres, chips, energy, and water resources to sustain increasingly capable models.

High performance comes at higher cost

That Anthropic PBC has sharply increased prices for premium Claude services used by finance professionals signals that while basic AI may eventually get commoditised, high-performance AI for finance, law, software engineering, and enterprise workflows may end up resembling premium infrastructure rather than cheap digital labour.

AI, thus, may not eliminate the cost of human work so much as redistribute it, replacing payroll expenses with infrastructure, supervision, and the potentially enormous cost of machine error.

The importance of being human

Human skills remain the most sought-after capabilities among employers globally, yet shortages in these skills are as severe as—and often more severe than—gaps in technical expertise, according to the QS Global Employer Survey. The QS World Future Skills Index 2027 finds that higher education systems are struggling to keep pace with a labour market being reshaped by AI.

The report argues that bridging this gap will require more than curriculum reform, calling instead for a broader reset of higher education policy, funding, governance, institutional agility, and employer investment in lifelong learning.

India highlights both the challenge and the opportunity. It ranks fifth globally in Future of Work readiness (96.0) but only 18th in Skills Alignment (82.7), underscoring the mismatch between labour market transformation and the ability to produce job-ready graduates. As one of the world’s largest higher education systems, India’s success in closing this gap will shape not only its own economic future but also the global supply of skilled talent, the report says.

Additional Reading

• Jeff Bezos rejects fears of job loss due to AI

• AI job apocalypse ‘unlikely,’ Goldman Sachs says

AI BITS & BYTES

Does RBI’s AI framework account for Agentic AI?

The Reserve Bank of India’s recently released draft guidance on an AI framework for banks deserves credit for addressing limitations ranging from hallucinations to model drift and adversarial attacks.The key proposals include strict validation for third-party models and mandatory “kill switch” mechanisms for AI.

That said, the next generation of AI systems may not behave like models at all. The Agentic AI systems can plan, make decisions, interact with external tools, execute actions and coordinate with other agents with limited human intervention.

In the context of banking, AI Agents may eventually approve loans, execute financial transactions, communicate with customers and trigger downstream actions autonomously. At that point, the question facing regulators changes fundamentally. The challenge is no longer supervising models that make predictions, but governing systems that make decisions and act on them.

Anthropic restores Claude Mythos, Fable

On Tuesday, Anthropic said that the US Commerce Department has lifted export controls on its most advanced AI models—Fable and Mythos. “We’ve received notice that the Department of Commerce has lifted export controls on Claude Fable 5 and Mythos 5,” Anthropic said in a post on X.

Anthropic had abruptly disabled its Mythos 5 and Fable 5 models following the export-control order on 12 June, over national security risks.

“We’ll begin restoring access tomorrow, and will share an update soon,” the company said. Anthropic also thanked users for their support. Read more.

Senior leadership gets new AI scorecards

Employers investing in AI-led upskilling programmes for their staff want their money’s worth. And the sure-shot way they are doing it is by including AI adoption as a parameter in performance appraisals.

From banking to consulting, business houses have started rating 5-15% of their senior leadership based on how they used AI beneficially, along with factors like people management, operations and financials. Read more.

Why an AI correction can be India’s opportunity

Big Tech remains publicly confident that artificial intelligence (AI) will eventually justify the huge sums of capital pouring into the sector. Yet, concerns have grown over less-than-commensurate returns and high market volatility amid reports that OpenAI may defer its long-awaited public offer of equity.

All this suggests that the breakneck pace of today’s AI boom may not be sustainable. Goldman Sachs projects capital spending of $5.3 trillion by hyperscalers between 2025 and 2030, much of it in AI chips and data centres.

Should today’s AI narrative falter or investment plans go awry, though, it may prompt investors to look anew at AI latecomers like India. Read more.

Hope you folks have a great weekend, and your feedback will be much appreciated—just reply to this mail, and I’ll respond.

Edited by Rashmi Sanyal.

Are these terms reflecting genuine shifts in how AI systems are being built and deployed? What exactly do these new roles and concepts mean, and why are they attracting so much attention?

In this week’s edition of Tech Talk:

• Meta gets India cred with Kunal Shah as WhatsApp boss

• Big Tech is panicking about AI and terrifying workers

• AI Tool of the Week: ChatGPT Scheduled Tasks

PREVIOUSLY, ON TECH TALK

Subscribe now

Anthropic PBC co-founder and Claude Code creator Boris Cherny, who had previously proclaimed that “software engineering is dead”, now asserts that the era of manual AI prompting is drawing to a close.

• Prompt engineering focuses on crafting the right instructions to get the best possible response from an AI model.

• Loop engineering is about designing workflows in which AI systems repeatedly reason, act, check their work and improve their outputs.

For example, a prompt engineer might write a prompt asking an AI to summarise a legal document. A loop engineer would build a system that reads the document, identifies ambiguities, searches relevant regulations, reviews its own conclusions and flags uncertain cases for human review. As AI becomes more agentic, the challenge shifts from writing better prompts to designing better processes, making loop engineering a natural evolution of the role.

Loop engineering is about encoding expertise. The best loop engineers are often not the people who know the most about AI. They are the people who understand a domain such as media, law, medicine, finance, or software development, and can translate that expertise into a repeatable sequence of checks, decisions and actions.

• Prompt engineering: You ask a question. The AI model answers.

• Context engineering: You provide the right documents, examples, memories, and tools for the model to deliver a better outcome.

• Loop engineering: You define how the model plans, acts, checks its work, uses tools, recovers from errors, and decides when the job is done.

These stages coexist. A coding agent still relies on good prompts. A research assistant still benefits from careful context management. Loop engineering simply becomes more important as you move from “generate an answer” to “complete an objective”.

This video by Anthropic engineers is instructive. DeepLearning.AI founder and Coursera co-founder Andrew Ng, says: “100% of my tasks are now done by AI agents—hype has exceeded my expectations. Loops is the next step. In 3-6 months, everyone will be using self-improving loops. No more prompting.”

That said, loop engineering can be significantly more expensive than prompt engineering because it typically requires multiple model interactions rather than a single request and response. A simple prompt-based task might involve one model call consuming a few thousand tokens. A loop-based system, however, may ask the model to plan, retrieve information, evaluate its findings, revise its output, verify its conclusions and, in some cases, seek human review before producing a final answer.

Each of these steps consumes additional tokens—meaning, a task that might require 5,000 tokens as a prompt could easily consume 25,000–100,000 when implemented as a loop. The higher cost comes from repeated reasoning, the accumulation of context across multiple steps and the need to process information returned by tools such as search engines, databases or code interpreters.

The trade-off is improved reliability

A single prompt may be sufficient to summarise a document, but a loop can check for errors, verify facts and revisit uncertain conclusions before responding.

As a result, loops are particularly valuable for high-stakes applications like software development, legal analysis and financial decision-making, where accuracy matters more than efficiency. Yet, as models become cheaper, token costs are likely to fall faster than labour costs. The cost of AI inference has declined dramatically while model capabilities have improved. If that trend continues, the economic penalty for using loops will become less significant over time.

Forward Deployed Engineer too is not an entirely new term, but AI has turned it from a niche role into one of the hottest jobs in tech. The term became widely associated with Palantir Technologies Inc., where engineers were sent directly to customer sites to understand real-world problems and build solutions alongside users rather than from a distant headquarters. Today, AI companies such as OpenAI, Anthropic, and Cursor have embraced similar roles.

An FDE is a hybrid of software engineer, product manager, consultant and AI workflow designer. Unlike a traditional software engineer deployed onsite by an IT services firm—whose role is usually to implement predefined requirements, integrate systems or maintain software—an FDE is expected to identify the problem, redesign workflows and determine how AI can create measurable business value. The focus is less on coding features and more on deploying intelligence.

As AI shifts attention from prompts to agents and business processes, FDEs are increasingly designing Human-AI workflows, evaluation loops and organisational change, making them an AI-native evolution of customer-facing engineering roles.

Meta can buy talent, but success takes a lot more

Lately, Meta Platforms Inc. CEO Mark Zuckerberg has been reaching for his cheque book ever so often to solve strategic problems.

Confronted with mounting competition in AI, for instance, Meta spent billions on talent. It recruited high-profile researchers, founders and entire teams from startups such as Dreamer, Maltbook and Scale AI (whose co-founder Alexandr Wang was asked to lead its quest for super-intelligence).

Now, Zuckerberg appears to have taken a similar playbook to WhatsApp. Meta has invested $900 million in Indian fintech startup Cred for a minority stake and hired its founder Kunal Shah to lead the chat platform globally.

India is central to Meta’s growth story. Of WhatsApp’s 3.3 billion monthly active users, a large chunk are Indians. Yet, for all its popularity, WhatsApp has been unable to grab ad revenue and has struggled to evolve into the financial powerhouse that Meta had envisioned it to be. Enter Kunal Shah—a serial entrepreneur with success in payments, credit and other financial services. Clearly, Meta wants to monetise WhatsApp quickly.

Recruiting startup founders can infuse fresh ideas and entrepreneurial energy, but that in itself does not create successful services. In other words, although Meta has shown that it can attract top talent, how it translates this into sustainable business growth is the actual test it faces. Here’s the full article.

AI BITS & BYTES

Most AI-led mainframe migrations will fall short: Gartner

AI is not a silver bullet for mainframe migration, and many companies may be better off modernising existing systems than replacing them, suggests a new Gartner report. More than 70% of mainframe migration projects launched in 2026 will fail to deliver their expected benefits because companies are overestimating what Generative AI can do, according to the research firm.

Gartner cautions that many organisations are being swayed by vendor claims that AI can quickly and easily modernise decades-old systems. In reality, migrating complex legacy applications remains difficult, risky and expensive.

ALSO READ | The Sword of Damocles that’s Anthropic Claude

As Claude Mythos falls, OpenAI GPT-5.5 Cyber rises

Last month, OpenAI granted the US government early access to GPT-5.5 for national security testing and evaluation. The latest frontier model is designed to tackle complex real-world tasks. On Monday, OpenAI expanded access to a cyber-focused version, GPT-5.5 Cyber, through a limited release to what it calls “trusted defenders”.

To that end, the company unveiled Daybreak, a platform that combines its frontier cyber models, Trusted Access for Cyber, Codex Security workflows, and third-party partners. The goal is to help approved defenders identify vulnerabilities, prioritise risks, generate fixes, and produce evidence within existing security and software-development workflows.

ALSO READ | Who gets to use Frontier AI?

Big Tech is panicking about AI and terrifying workers

A three-decade veteran of Microsoft Corp. is pushing back against what he describes as a culture of manufactured alarm inside the technology industry. Brad Smith, who serves as vice-chairman and president of the company, says influential voices in tech have been consistently wrong about the pace of disruption, consistently hypocritical about the solutions they propose, and consistently careless about the human cost of the panic they generate.

ALSO READ | We can’t let AI giants eat the economy: Nadella

Hope you folks have a great weekend, and your feedback will be much appreciated—just reply to this mail, and I’ll respond.

Edited by Rashmi Sanyal.

Thousands of fans have travelled across continents to support their national teams at the 2026 FIFA World Cup in North America. Millions more are glued to screens of all shapes and sizes, watching 22 players chase a ball across a field.

The tournament’s global appeal has turned sport into a formidable economic force. The FIFA World Cup, the world’s most-watched sporting event, could generate more than $40 billion in global GDP, according to the World Economic Forum. In its January report, Sports for People and Planet, produced with Oliver Wyman, WEF estimated that the global sports industry could grow into an $8.8 trillion economy by 2050.

Yet, even as billions follow the World Cup, advances in AI are producing humanoid robots capable of tasks once considered uniquely human—from running races and playing table tennis to kicking a football. In laboratories and testing grounds from Beijing to the US, intelligent machines are competing in driverless racing leagues and robot tournaments such as RoboCup and the World Robot Olympiad.

In this week’s edition of Tech Talk:

• AI tool of the week: ZoomMate

• Who gets to use frontier AI?

• ChatGPT cedes ground to Claude, Gemini

PREVIOUSLY, ON TECH TALK

Subscribe now

AI in Sports

AI has long outperformed humans in Chess and Go, games once regarded as the ultimate tests of intellect. But while computers excel at solving math problems or optimising routes, they continue to struggle with physical tasks that humans perform effortlessly, from picking up a cup to making a chapati. This gap between what humans find easy and machines find difficult is known as Moravec’s Paradox.

Football offers a vivid example. Even a child can kick and dribble a ball with ease. For a robot, however, the same action requires a complex sequence of calculations and movements. It must locate the ball, move into position, maintain balance on one leg and coordinate the other to strike with the required force and accuracy. Replicating a footballer’s abilities in a humanoid robot demands the seamless integration of perception, locomotion, balance and motor control.

This month, Hyundai Motor Co. and Boston Dynamics showcased these capabilities through football drills performed by Atlas, including “Ghost Rabona”—a feint followed by a cross-legged kick. Researchers first captured the movements of professional players using optical motion-capture systems and adapted them to Atlas through a process called retargeting, which translates human actions to the robot’s body geometry and joint constraints.

Atlas then refined these skills using reinforcement learning in large-scale simulations running on cloud GPUs, practising thousands of times in parallel before the resulting control policies were deployed on the physical robot.

Such demonstrations are becoming increasingly common.

China’s investment in humanoid robotics has produced robots that run races and participate in sporting events, while robot combat competitions and autonomous racing leagues have found audiences of their own.

The RoboCup initiative, launched in the late 1990s, aims to develop a team of fully autonomous humanoid robots capable of defeating the reigning FIFA World Cup champions by 2050. Even robot competitions increasingly mirror human sporting traditions, complete with national teams, medals and cheering spectators. Yet, as machines become more capable, the appeal of human competition remains undiminished, with the FIFA World Cup being a case in point.

Human vs AI in Sports

If the games we play were simply a quest to identify the fastest, strongest or most efficient performer, many of its defining traditions would make little sense. Formula One would become a competition among autonomous vehicles. Olympic sprinting would embrace exoskeletons and performance-enhancing technologies. Football teams would be free to engineer athletes biologically in pursuit of victory.

But sport has never worked that way. When IBM’s Deep Blue defeated Garry Kasparov in 1997, many predicted the decline of chess. Instead, the game endured. Likewise, AlphaGo’s victory over Lee Sedol transformed Go without diminishing the appeal of human competition. In both cases, people continued to value human achievement within human constraints.

Sport is not merely about maximising performance. It is about defining the conditions under which achievement acquires meaning. The 100-metre sprint is compelling not because it reveals the fastest possible movement on Earth—a cheetah would comfortably outrun Usain Bolt—but because it shows what a human being can accomplish within agreed rules and constraints.

Football captivates audiences for much the same reason. It showcases not only physical skill but also strategic intelligence, teamwork, resilience and the ability to perform under pressure. Fans are drawn not simply to the outcome, but to the human drama behind it. That is why moments such as Lionel Messi’s hattrick for Argentina at the World Cup resonate so deeply—they represent the limits and possibilities of human performance, not the optimisation of a machine.

While we may agree or disagree with these rankings, behind every human champion lies years of training and an ecosystem of coaches, analysts and support staff. Robotic competitions, in contrast, are mostly about engineering ingenuity and tech progress. Their protagonists are often the scientists and programmers behind the machines. A robot missing a decisive penalty is a technical failure; a human missing one can become tragedy, redemption or national folklore.

Machines may eventually surpass humans in sport. Yet the World Cup is unlikely to lose its appeal. Robot competitions may show how far technology has advanced. Human sport reveals who we are—and what we can become.

Controlling Intelligence: Who gets to use Frontier AI?

Four years ago, US chip export controls sought to deny the likes of China the means to build advanced AI systems. Last week’s curbs placed by the Trump administration on access to Anthropic’s frontier models Fable 5 and Mythos 5 by foreign nationals, including researchers at American labs, suggests that AI models themselves are now being treated as strategic assets.

If the first phase of US policy intervention in the AI race was about control of the chips that power AI, the next stage is about controlling who gets to use that intelligence. The contest, in other words, is moving up the stack—from chips to infrastructure, and now to AI itself.

Further, if comparable systems from OpenAI, Meta and Google remain unaffected, US policymakers will be asked to explain why. Without such clarity, legitimate AI safety precautions would be hard to disentangle from industrial policy and geopolitics.

That said, a precedent has been set. Countries can no longer assume that access to frontier intelligence will remain frictionless or politically neutral. This strengthens the case for sovereign AI ecosystems that encompass compute infrastructure, local datasets and foundational models. You may read the full piece here.

AI BITS & BYTES

Sarvam AI raises $234 million from HCLTech, others

Sarvam AI, India’s full-stack Sovereign AI startup, has raised $234 million as part of a targeted broader funding round as it looks to roll out AI models that can challenge Chinese and US rivals. HCLTech and venture capital firm Bessemer Venture Partners invested in the funding round, which also included existing investors such as Peak XV Partners and Khosla Ventures. Read more.

India’s IT industry will be gone: Vinod Khosla

India’s $200-billion IT services and BPO industry could face a major disruption from AI unless companies rapidly adapt their business models, according to Vinod Khosla, Silicon Valley venture capitalist and early-stage investor in OpenAI—the world’s most valuable AI startup. Read more.

ChatGPT’s market share falls below 50% for first time

Anthropic's Claude earns an average of $2.76 per user, 1.5 times what ChatGPT generates, and converts a higher share of its iOS base into paying subscribers, according to new Sensor Tower data, as OpenAI's grip on AI application market weakens significantly. Read more.

ALSO READS

• The AI challenge for India's middle managers

• DeepSeek is now China’s most valuable AI startup

• House robots are coming—and they will be dangerously cute

• Elon Musk is unleashing SpaceX to solve his AI problem

• Meta's own engineers are still reaching for Claude

Hope you folks have a great weekend, and your feedback will be much appreciated—just reply to this mail, and I’ll respond.

Edited by Rashmi Sanyal.

The reported move comes as OpenAI has confidentially filed an S-1 with the US Securities and Exchange Commission, formally taking its first step toward a public listing. SpaceX Inc. and Anthropic have also filed S-1s, while Perplexity has publicly disclosed its IPO ambitions.

In this week’s edition of Tech Talk:

• AI IPOs: SpaceX, OpenAI & Anthropic

• Fable 5, Claude Mythos in a safe avatar

• AI tool of the week: Gemini Daily Brief

• Anthropic says AI may eventually build itself

• TCS aims to have as many AI agents as humans

PREVIOUSLY, ON TECH TALK

Subscribe now

A ‘super app’, and an IPO

OpenAI, valued at roughly $850 billion and eyeing a listing, is burning through cash faster than it earns. Sam Altman may need such clever rebranding to push sales by transforming ChatGPT from a conversational chatbot into a full-stack “super app”. A senior OpenAI employee recently told the Financial Times that “chat is dead”. What the company needs is a platform that does things: books flights, writes code, designs presentations, places orders—all without the user leaving the ChatGPT interface.

To that end, OpenAI has unveiled an Apps SDK that embeds third-party services like Booking.com, Canva, Spotify, and Expedia directly into ChatGPT’s chat window. Codex, its AI coding assistant with over five million weekly active users, is also being pushed to the foreground. Autonomous agents capable of executing multi-step tasks are being positioned as the product’s new centre of gravity.

OpenAI needs to show Wall Street a credible path to profitability before it goes public. A super app, if it works, would deepen user engagement, justify premium enterprise subscriptions, and extract a toll from every third-party transaction conducted on its platform. The ambition, in short, is to become the operating system of the AI era.

History, however, counsels caution. Web portals tried precisely this in the late 1990s. Yahoo bundled search, email, news, finance, maps and social features into a single destination—and was undone not by a rival bundle, but by Google’s ruthless focus on doing one thing better than anyone else. When a superior alternative exists for any individual service within a bundle, users defect.

The Chinese super app experience is another case in point. WeChat succeeded because it controlled payments, owned the social graph, and operated within a relatively closed mobile ecosystem that favoured incumbency. Baidu attempted something similar and was steadily outflanked by rivals who controlled distribution channels it could not access.

OpenAI enjoys none of WeChat’s structural advantages. Apple and Google control the mobile OS layer in its key markets. European regulators have already restricted its third-party integrations across the UK, EEA and Switzerland with no committed timeline for resolution.

The early commercial evidence, too, is sobering. When Walmart piloted Instant Checkout inside ChatGPT, conversion rates fell well below its own channels. Walmart promptly replaced it with a proprietary assistant embedded in ChatGPT, retaining customer data and the transaction relationship while using OpenAI merely for distribution.

That said, what distinguishes OpenAI’s bet from its predecessors is the genuine novelty of AI agency—the ability to act across the internet on a user’s behalf, not merely retrieve and display. If that agentic capability proves transformative enough, it may generate the loyalty that bundling strategies alone never could. That remains, for now, an empirical wager rather than a proven proposition. The super app is the oldest trick in the technology playbook. OpenAI is staking its valuation on AI to make it work.

AI IPOs: A Deja Vu for investors

Investors have been here before. A transformative technology arrives, valuations detach from earnings and markets start pricing the future today. Capital markets financed booms in railways, dotcoms, social media and crypto plays, only to watch many companies collapse once reality crept up on what former Fed Chair Alan Greenspan called “irrational exuberance”.

Investors are now being asked to bankroll a new crop of companies whose sky-high valuations depend less on what they earn today than on what they might achieve someday. Think of the interplanetary ambitions of SpaceX or the quest for AGI at OpenAI and Anthropic.

With IPOs lined up in the US, those three American companies are expected to seek $220-250 billion from investors. Since only thin slices of equity will be listed on the stock market for trading, their valuations would be eye-popping: SpaceX worth almost $1.8 trillion, OpenAI and Anthropic at $1 trillion each. A sudden expansion in market capitalisation of such magnitude, even if only a fraction is free-float, would need a lot by way of business prospects to justify.

Here’s the full article.

AI BITS & BYTES

Fable 5, Claude Mythos in a safe avatar

Anthropic PBC has unveiled Fable 5, a new Mythos-class AI model that it says is safe for general use, while outperforming all previous Claude models such as Opus, as well as rival models of OpenAI and Google Gemini on multiple benchmarks.

The Claude maker has also introduced Mythos 5, a version of the same underlying model with certain safeguards relaxed for a limited group of cyber defenders and infrastructure providers through Project Glasswing.

Still, Anthropic warns that these AI models could significantly lower the cost and complexity of cyberattacks through advanced vulnerability discovery and agentic hacking capabilities. Read more.

Anthropic warns AI may eventually build itself

AI is likely to eventually reach a stage where it build itself—a concept known as recursive self-improvement, Anthropic noted in a blog.

“We are not there yet, and recursive self-improvement is not inevitable. But it could come sooner than most institutions are prepared for.”

The length of tasks that AI models can reliably complete on their own is doubling roughly every four months, versus seven months earlier. If this trend holds, tasks that take a skilled person days could come into range this year. In 2027, AI systems could be capable of tasks that take a person weeks.

Incidentally, more than 80% of the code merged into Anthropic’s codebase was authored by Claude as of May. Read more.

TCS wants as many AI agents as human workers

Tata Consultancy Services Ltd. wants to become the world’s largest AI services firm—a goal that would require as much artificial intelligence as human. At least that’s what Chairman Natarajan Chandrasekaran would have us believe.

A slowdown in business—TCS’s annual revenue shrunk for the first time since listing in FY26—is a thing of the past, Chandra said at the company’s thirty-first annual general meeting, outlining a future where the IT bellwether could deploy as many AI agents as employees, and embed AI in all revenue streams.

That, from a company that undertook its biggest layoffs last year. TCS ended FY26 with 584,519 employees, down by 23,460 compared to the previous year.

“Will AI lead to decrease in hiring? Absolutely. The company will not be hiring as it used to, Chandra said, “because certain portions of the work…will go to agents.”

🔗 The Sword of Damocles that’s Anthropic Claude

Meta + Reliance’s AI data centre in India

Meta Platforms Inc. has partnered with Reliance Industries Ltd. to lease its first AI-enabled data centre in India. RIL will build a 168 MW capacity data centre in Gujarat’s Jamnagar, which will then be leased by Meta to power its products and AI. That follows Google’s plans to set up a $15 billion, 1 GW AI hub in Andhra Pradesh’s Visakhapatnam.

According to IMARC Group, India’s data centre market is projected to nearly double to $13.11 billion by 2034, fuelled by increasing cloud adoption, digital transformation initiatives, and rising demand for AI computing capacity.

IN CHARTS

In May 2026, global cyberattack activity eased from April’s sharp rebound, but organisations experienced an average of 2,055 weekly cyberattacks—a 2% increase annually and a 7% decrease monthly.

While the monthly decline may read as stabilisation, ransomware activity surged to its highest year-on-year growth rate of 2026, and Gen AI-driven data exposure risks continued to deepen across enterprise environments.

ALSO READS

• Siri AI to iOS 27: Every new feature Apple announced at WWDC 2026

• Pentagon blacklists Baidu, BYD weeks after Trump-Xi summit

• Here’s how long it will take for AI to reach its potential

• Does workplace loyalty still matter when AI can take your job?

• The AI revolution worth betting on isn’t in your portfolio

Hope you folks have a great weekend, and your feedback will be much appreciated—just reply to this mail, and I’ll respond.

Edited by Rashmi Sanyal.

Nvidia CEO Jensen Huang believes these RTX Spark-powered Windows machines will transform the PC from a passive tool into an active digital teammate, capable of understanding instructions and carrying out tasks on a user’s behalf.

The logic is simple. Everyday activities such as word processing, text messaging, web browsing and basic photo or video editing can easily run on conventional operating systems and low-power processors. AI models, however, demand greater computing power, memory and software optimised for AI workloads. That is the future Nvidia is betting on with its new line-up of desktops, laptops and workstations powered by its RTX Spark “superchip”.

Nvidia says the systems can run some of today’s most advanced AI models locally—workloads that would typically require cloud-based computing resources. They are aimed at AI content creation, software development, professional video-editing and high-end gaming, while retaining the portability expected of modern laptops.

The headline specification is 1 petaflop of AI performance. A petaflop equals one quadrillion, or 1,000 trillion, calculations per second.

As AI models grow larger and more sophisticated, they require exponentially more computing power. For perspective, the computers that helped land humans on the moon had less computing power than today’s smartphone. Likewise, petaflop-scale computing was once the preserve of supercomputers used in scientific research, weather forecasting and defence. Nvidia is now attempting to bring that class of AI computing to a laptop.

In this week’s edition of Tech Talk:

• Anthropic confidentially files for IPO

• Sam Altman changes his mind on AI layoffs

• AI Tool of the Week: ChatGPT for PowerPoint

• AI Lexicon: What is Tokenmaxxing?

PREVIOUSLY, ON TECH TALK

Subscribe now

An Apollo moment for consumers?

Speaking at Nvidia GTC Taipei 2026, Huang painted a future in which AI becomes as commonplace as televisions and dishwashers, with AI supercomputers running personal agents in homes and, eventually, powering everything from robots and factories to satellites and telecommunications networks.

It is an ambitious vision, but has gaps to fill. Smartphones transformed computing because they evolved to combine communication, internet connectivity, navigation, photography, entertainment, shopping, and even banking in a single device. But unlike smartphones, which solved obvious consumer problems at the outset, Agentic AI PCs have yet to demonstrate why most people need supercomputers on their desks.

Moreover, the first wave of AI PCs already promised on-device AI through neural processing units, local copilots, image generation, transcription, and productivity features.

Yet, most users still spend their time in browsers, MS Office apps, streaming services, messaging apps, and games. Very few people are running local 70B-120B parameter models, agent swarms, or autonomous workflow systems.

So an average buyer may reasonably ask: What can an Agentic AI PC do that an AI PC equipped with an NPU, GPU and access to cloud-based AI services cannot?

That is why the most immediate audience for Agentic AI PCs is likely to be developers, researchers, robotics, data scientists, quantitative analysts, advanced creators, AI startups and enterprise teams. Such users already spend significant amounts on cloud computing and can justify investments in powerful local hardware.

Moreover, for them, running large models locally offers lower inference costs, better privacy, reduced cloud dependence, faster experimentation, and the ability to build persistent agents.

From ‘Intel inside’ to ‘Nvidia inside’?

The launch is not just about a new category of PC. It is also part of a battle over who owns the next computing stack. For more than four decades, personal computing revolved around the central processing unit, or CPU. Intel Corp. became one of the most influential companies in technology because CPU performance was the metric by which computers were judged. Nvidia’s ambitions suggest a different future. If AI workloads become the defining function of computing, then AI performance could become more important than traditional processor performance.

Nvidia gains influence because its expertise lies in AI acceleration. AMD benefits because it competes across both CPUs and GPUs. Qualcomm and ARM-based designs could become increasingly relevant if always-on, energy-efficient AI proves more important than raw processing power.

Intel, meanwhile, faces the uncomfortable prospect of seeing the industry’s centre of gravity move away from the category it has dominated for decades. Seen this way, the battle is not really about Agentic AI PCs. It is no longer about: “Which processor is inside?” but “Where does AI happen?”

“Agentic computing” also represents a shift from app-centric interfaces to outcome-driven interactions. The idea of delegating tasks rather than manually navigating software has obvious appeal. But history suggests that consumers adopt new technologies because they remove friction, not because they introduce impressive capabilities. Agentic computing may well become the next chapter in personal computing, but only if it serves consumer needs.

Twist in the tale

There’s another possibility we cannot rule out. Agentic AI may ultimately evolve into a cloud service rather than a hardware category. In that scenario, companies such as OpenAI, Anthropic, Google and Microsoft would run the agents in their data centres, while users access them from relatively simple devices. The shift would mirror earlier transitions in computing, where workloads that once required powerful desktops gradually move to browsers, SaaS applications and cloud infrastructure.

If Agentic AI follows the same path, hardware becomes less important and the differentiation between devices becomes harder—a prospect that would challenge not only Intel, AMD and Qualcomm but Nvidia itself, and the future of Agentic AI PCs.

AI BITS & BYTES

Anthropic takes first step toward IPO

Anthropic PBC has confidentially filed draft paperwork with the US Securities and Exchange Commission for a potential initial public offering, taking the first formal step toward a listing. The company has not disclosed how many shares it plans to sell or at what price.

The filing came days after the Claude maker announced a $65 billion Series H funding round led by Altimeter Capital, Dragoneer, Greenoaks and Sequoia Capital, valuing the company at $965 billion. Anthropic also said that its annualised revenue run rate crossed $47 billion in early May, underscoring the rapid growth of the AI industry.

For comparison, OpenAI’s most recent funding round closed on 31 March with $122 billion in committed capital at an $852 billion post-money valuation. At the time, the ChatGPT maker said it was generating roughly $2 billion in monthly revenue, equivalent to an annualised run rate of about $24 billion. Read more.

Anthropic launches Claude Opus 4.8

Anthropic has introduced Claude Opus 4.8, with improvements across benchmarks to become a more effective collaborator. It’s available today for the same price. The table below shows how Opus 4.8 compares to its predecessor and other models on tests of coding, agentic skills, reasoning, and practical knowledge work tasks.

India gets access to Claude Mythos

Anthropic is finally expanding access to its powerful Mythos AI model via Project Glasswing. In a blogpost on Tuesday, the company announced that it will be bringing Mythos to 150 additional organisations across 15 countries, including India. Other nations include Canada, Australia, New Zealand, France, Germany, Italy, Switzerland, the Netherlands, Spain, Belgium, Sweden, Japan, and South Korea. Read more.

ALSO READS

• Alphabet to raise $80 billion from stock sale

• AI isn’t causing mass layoffs, Sam Altman says in U-turn

• Cheaper AI could create more jobs, not fewer: Apollo’s Torsten Sløk

• Cloud war intensifies as Google ties enterprise deals to in-house AI

• Microsoft faces backlash as Github Copilot ends flat-rate AI pricing

Hope you folks have a great weekend, and your feedback will be much appreciated—just reply to this mail, and I’ll respond.

Edited by Rashmi Sanyal. Produced by Tushar Deep Singh.

Most recently, Meta Platforms Inc. laid off nearly 8,000 employees globally as part of Chief Executive Officer Mark Zuckerberg’s broader AI-driven restructuring strategy, while LinkedIn announced more than 600 job cuts set to take effect in mid-July. According to layoffs.fyi, 152 tech firms have cut 115,430 jobs so far in 2026.

That number is likely to rise as Generative AI increasingly writes reports, generates code, analyses data, creates presentations and mimics human conversation with startling fluency. AI agents are pushing automation even further by planning, reasoning and carrying out multi-step tasks with minimal supervision.

Optimists argue that, like previous technological revolutions, AI will augment workers, boost productivity, and create entirely new jobs and industries. For instance, Sam Altman believes AI will not trigger a global “jobs apocalypse”. Pessimists counter that this wave is fundamentally different because AI is scaling cognitive capability at an accelerated pace and at near-zero marginal cost.

Speaking at a Vatican gathering where Pope Leo XIV formally released his first AI-focused encyclical, Magnifica Humanitas, and called for AI to be “disarmed” to protect humanity from its unchecked dangers, Anthropic PBC co-founder Christopher Olah warned that AI could trigger “large-scale” job displacement, echoing co-founder and CEO Dario Amodei’s warning.

In this week’s edition of Tech Talk:

• Your boss may be quietly sharing your data with Big Tech

• Flipkart bets AI can turn internet users into shoppers

• Seven AI gadgets of 2026 that actually feel useful

• AI tool of the week: OpenAI’s Verify

PREVIOUSLY, ON TECH TALK

Subscribe now

White-collar to blue-collar

Anxiety over AI is no longer confined to white-collar work. There is growing concern that AI robots can increasingly disrupt physical labour as well. In May 2025, Bank of America Global Research predicted that humanoid robots will see mass adoption in highly unstructured environments—from households to elderly care—by 2035. The firm forecast annual humanoid shipments could reach 1 million units between 2030 and 2035, before eventually rising to 3 billion globally by 2060.

Those projections rest on bold assumptions: humanoids replacing 20% of industrial workers and 50% of service-sector workers; each robot substituting for 2.5 industrial workers or 1.5 service workers; and household penetration eventually reaching 0.7 robots per home.

Whether those estimates materialise remains uncertain. What is clear, however, is that humanoids are becoming increasingly capable as Gen AI and Agentic AI are fused into robotic systems, giving machines the ability not just to execute commands, but also to reason, adapt and autonomously complete tasks.

The trajectory is striking. Back in 2017, Sophia—the humanoid developed by Hanson Robotics—became the world’s first robot to receive citizenship from Saudi Arabia. This month, Unitree Robotics unveiled the GD01, a nearly 10-foot-tall mecha capable of shifting between bipedal and quadrupedal movement, evoking something closer to the movie ‘Transformers’ than traditional industrial machinery.

Attention is all you need

China is reportedly planning to assign digital identities to every humanoid robot manufactured in the country, enabling lifecycle tracking from production to recycling. According to reports, more than 100 Chinese humanoid manufacturers and over 28,000 robots across 200 models have already been assigned such IDs.

Yet the larger question is no longer whether robots will enter workplaces, but how deeply they will reshape labour itself.

Amazon.com Inc. offers perhaps the clearest glimpse of that future. Since acquiring Kiva Systems in 2012 and launching Amazon Robotics, the company has deployed more than one million robots across its operations network. The company now integrating Gen AI and Agentic AI into its robotic ecosystem to create increasingly autonomous machine labour. Its DeepFleet Gen AI foundation model is designed to make the company’s robotic fleet smarter and more efficient. Project Eluna uses agentic AI to optimise workflows and improve operational safety.

The machines themselves are also becoming more sophisticated. Hercules robots can lift and transport up to 566 kg of inventory. Pegasus uses precision conveyor systems to sort individual packages. Proteus, Amazon’s first fully autonomous mobile robot, can navigate around workers in open environments while hauling heavy carts. Vulcan adds another layer of capability with a sense of touch, allowing it to detect and adjust contact with objects.

Costs: Robotic labour vs Human labour

For years, humanoid robots were sold as the ultimate cheap workforce—tireless, scalable, and immune to wages, unions, or burnout.

But a single advanced robot may contain thousands of components and hundreds of moving mechanical parts—actuators, motors, gearboxes, sensors and cameras. Many systems also require expensive graphics processing units and continuous AI inference to operate in real time. That hardware stack is costly. The likes of Unitree Robotics, Boston Dynamics, Tesla Inc. and Figure AI are pushing prices down, but even then, commercial humanoids can cost anywhere between $20,000 and $150,000, or more.

And the robot itself is only part of the bill. Deployment also involves software integration, safety systems, cloud infrastructure, maintenance, remote monitoring, and human supervisors. In many cases, so-called autonomous systems still rely on tele-operators and engineers working quietly behind the scenes. The irony, hence, is that replacing labour currently requires enormous amounts of human labour.

Yet, the economics may change precisely because robots are starting to participate in manufacturing itself. Industrial robotic arms already build components used in other robots. Unitree Robotics is already experimenting with humanoids on assembly lines, while Boston Dynamics’ Atlas is being positioned for industrial logistics and manufacturing environments.

This creates a potential feedback loop: Humans build early robots. Robots help manufacture future robots. Production costs fall. Automation deepens. And even more robots become economically viable. While the vision of factories automating their own expansion is seductive, there is a massive gap between robot-assisted manufacturing and fully autonomous robotic self-replication. Humanoids remain poor at delicate wiring, troubleshooting, repairs, and handling chaotic real-world variability.

For now, human labour is cheaper, more adaptable, and easier to deploy. A warehouse worker can improvise, recover from mistakes, navigate uncertainty, and perform multiple physical tasks without needing millions of dollars in training data or precision engineering. A humanoid robot, by contrast, may struggle with something as simple as a loose cable or an unexpected object placement.

You may also read about the disillusionment with AI software tools that are proving more expensive than human labour: ‘AI promised cost savings, but Microsoft and Uber say it’s costing more than human workers’.

Economics depends on Geography

In high-wage economies facing labour shortages and aging populations, robotic labour may eventually make sense. In lower-wage environments like India, the numbers are far less compelling.

That said, like early computers or electric vehicles, humanoids may currently appear inefficient, expensive, and overhyped until scale changes the equation. If hardware prices fall, AI models improve, and factories become increasingly automated, robotic labour could eventually become dramatically cheaper than human labour in repetitive industrial work.

But that future is not here yet. Today’s AI humanoids are less a replacement for human workers than a reminder of how much invisible human effort still underpins the automation revolution itself. But we need to prepare ourselves for a future where robots take over routine chores and tasks, similar to how washing machines, vacuum cleaners and food processors are complementing—and, in some cases, even replacing—house help.

🔗 Mint Explainer | Why Pronto’s physical AI training pilot is drawing scrutiny

AI BITS & BYTES

Your boss may be sharing your data with Big Tech

The app your employer uses to track attendance, productivity or work hours could also be sending your data to tech giants like Google, Meta and Microsoft. A new study has found that these workplace monitoring apps, often referred to as “bossware”, are also sharing employee data with digital advertising platforms and data brokers.

Not a fan of Google AI Search? Try these instead

As Google expands AI Summaries and conversational search, many users are exploring alternatives. Here are six search engines offering privacy, fewer ads and a cleaner web experience. You may also read: Google won’t let you disable AI in Search, but these tricks still work.

7 AI gadgets of 2026 that are actually useful

AI gadgets in 2026 are finally moving beyond hype, with smart glasses, AI rings, fitness trackers, and wearable assistants offering practical features. From Fitbit’s AI-powered wellness insights to Meta’s smart glasses and Samsung’s Galaxy Ring, these devices focus on convenience and productivity.

Hope you folks have a great weekend, and your feedback will be much appreciated— just reply to this mail, and I’ll respond.

Edited by Rashmi Sanyal. Produced by Tushar Deep Singh.

AI is marked by a similar lack of consensus on foundational ideas. What qualifies as reasoning? Is a model that predicts tokens actually “thinking”? Does intelligence require consciousness, or is it simply a matter of function and behaviour?

For instance, Artificial General Intelligence, or AGI, means different things to OpenAI, DeepMind, and Meta. “Reasoning models” are celebrated by some as a breakthrough and dismissed by others as sophisticated pattern-matching or stochastic parrots. “Consciousness” in AI is either imminent, impossible, or simply the wrong question, depending on who you ask—a believer, an AI researcher, scientist, or a philosopher.

In this week’s edition of Tech Talk:

• Meta, LinkedIn lay off hundreds amid AI shift

• Google I/O 2026: AI overhaul of Google Search

• AI Tool of the Week: Claude for Small Business

PREVIOUSLY, ON TECH TALK

Subscribe now

Broadly speaking, there are two camps. On one side are experts such as Fei-Fei Li (co-founder and CEO of World Labs, also known as the Godmother of AI), Yann LeCun (former Meta Platforms chief scientist and executive chairman of Advanced Machine Intelligence, and also one of the three godfathers of AI along with Geoffrey Hinton and Yoshua Bengio), and Andrew Ng (co-founder and head of Google Brain, former chief scientist at Baidu, and currently co-founder Coursera).

This camp argues that AI still has a long way to go before approaching human-like intelligence, that fears of mass job destruction are overstated, and that AGI remains, at least for now, an overreach.

On the other hand, experts like Geoffrey Hinton (Nobel Laureate), Dario Amodei (Anthropic CEO), Sam Altman (OpenAI CEO) and Mustafa Suleyman (Microsoft AI CEO) believe that AGI is just a couple of years away and could fundamentally reshape labour markets and institutions.

In essence, while the first camp sees today’s AI systems as pattern-matching tools still far from robust reasoning and autonomy, the second views current AI as an early form of AGI capable of rapid, exponential improvement.

Five broad AGI assumptions

Many debates about AGI quietly assume that machine intelligence and human intelligence are equivalent forms of cognition moving along the same curve. But today’s reasoning models, while extraordinarily capable, may still operate in ways fundamentally different from biological intelligence.

• First, human reasoning evolved through survival in the physical world—movement, pain, hunger, social interaction, mortality, sensory grounding. Our intelligence is inseparable from embodiment. AI models, by contrast, are mostly statistical abstractions trained on representations of the world, not direct lived experience of it.

• Second, even advanced reasoning models frequently infer from patterns rather than from deeply grounded causal world models. They can simulate causal reasoning impressively, but simulation is not necessarily understanding. A human child may understand physical causality from touching fire once; a model may require millions of examples. The danger is that fluency can masquerade as comprehension. Further, humans know—to some degree—what they don’t know. AI models are structured in such a way that they confabulate with the same fluent confidence—whether they are right or wrong.

• Third, large language models can discuss quantum physics, poetry, law and biology within seconds—something no human can do with equal breadth. Yet humans possess continuity of self, durable goals, lived memory, emotional salience and contextual grounding accumulated over years. AI resembles a vast probabilistic map; human intelligence resembles an adaptive organism.

• Fourth, AI does not need consciousness, self-awareness, or human-style understanding to outperform humans economically. Calculators surpassed humans in arithmetic without “understanding” mathematics. Aircraft surpassed birds without flapping wings. Similarly, AI may dominate many cognitive tasks without possessing anything resembling human consciousness or general intelligence.

• Fifth, the obsession with AGI itself is mostly anthropocentric. Humans often evaluate AI by asking: “Can it think or act like us?” But history suggests superior systems rarely mimic nature exactly. Submarines do not swim like fish; search engines do not remember like humans. AI may evolve into a non-human form of intelligence that is immensely capable yet permanently alien—powerful in areas humans value economically, weak in areas humans intuitively associate with “real intelligence”.

Hence, the disagreement between AI experts exists because they are not debating the speed of progress but they are also operating from different definitions of intelligence, different scientific intuitions, and different philosophical priors about what current AI systems actually are. The fact, though, is while today’s AI systems may indeed lack deep human-style understanding, they are becoming economically and strategically transformative at extraordinary speed, which is what makes the debate feel so polarised.

That said, whether AGI arrives in 2027 or 2047, the question is the same: who decides what it optimises for, who owns it, and who bears the cost when it gets things wrong? Society has largely been a spectator in this debate that is arguably one of the most consequential in human history. This need not be the case.

AI BITS & BYTES

Meta, LinkedIn let go hundreds of employees

Meta Platforms Inc. has started laying off nearly 8,000 employees globally, with workers in Singapore among the first affected. The cuts are part of CEO Mark Zuckerberg’s broader restructuring strategy as the company increases AI spending and streamlines operations. LinkedIn, too, has announced job cuts for over 600 employees which will take effect on 13 July. In 2026 so far, 111,173 tech employees have been laid off by 147 tech companies, as per layoffs.fyi.

🔗 Prepare for an AI jobs apocalypse

7 AI scams Indians are falling for in 2026

AI-powered scams are becoming more sophisticated in India, with fraudsters using cloned voices, deepfake videos and realistic phishing emails to target users. From fake HR interviews and “digital arrest” scams to fraudulent AI trading apps, cybercriminals are exploiting trust at scale. Here’s more on this topic.

Google I/O 2026: AI overhaul of Google Search

Google has introduced a new world model called Gemini Omni. The new model combines the reasoning abilities of other generative media models from Google including Veo, Nano Banana, and Genie. Google DeepMind CEO Demis Hassabis called Omni a significant leap toward AGI. You can read more about this here.

Hope you folks have a great weekend, and your feedback will be much appreciated— just reply to this mail, and I’ll respond.

Edited by Rashmi Sanyal. Produced by Tushar Deep Singh.

Consider San Francisco-based AI startup Andon Labs, where human baristas still brew and serve coffee, but a Google Gemini-powered AI agent named “Mona” oversees almost everything else at the company’s Andon Cafe in Sweden. According to an AP report, Mona handles tasks ranging from hiring staff to managing inventory.

While the cafe itself is reportedly struggling to turn a profit, the novelty has clearly resonated with customers. But will this curiosity evolve into sustained consumer interest? More importantly, could AI agents eventually run businesses autonomously? And if they do, what would that mean for human jobs, shareholders, and society at large?

In this week’s edition of Mint Tech Talk:

• India uses AI tools for weather forecasting

• Ilya Sutskever’s bombshell on OpenAI-Anthropic merger

• AI Tool of the Week: ChatGPT Memory Sources

• What machines say about future of human jobs

PREVIOUSLY, ON TECH TALK

Subscribe now

Founded in 2023, Andon Labs partnered with Anthropic last June on Project Vend—an experiment in which Claude Sonnet 3.7 operated a small automated store inside the AI startup’s San Francisco office. Nicknamed “Claudius”, the AI shopkeeper was tasked with handling complex commercial operations such as inventory management, pricing decisions, and avoiding bankruptcy over an extended period of time.

Startups such as Rocketable and ecosystems built around OpenClaw, too, suggest that the first versions of AI agents running companies are already emerging, not as fully autonomous corporations but as “agent-native” businesses designed around AI workers from day one. The result: A startup that once required dozens of employees may increasingly operate with one or two founders supervising fleets of AI agents.

For instance, Bridgewater’s AIA Labs uses AI agents to break down complex investment strategies, speeding up research while keeping analysts in the loop. The Artificial Investment Associate (AIA) serves as the primary decision-maker in the fund, while humans oversee risk management, data acquisition, and trade execution. Like wise, Sydney-based Minotaur Capital is an AI-powered hedge fund that has outperformed the market.

Former Google CEO Eric Schmidt believes that “if you really want to make money, it’s actually easy—found an agentic AI company. I mean, build an agent to do something. This is the agentic period in AI. For the next year or two, everyone is going to build agents. The agents are all going to compete”.

Investors too are excited about the prospect of “one-person unicorns”—tiny companies that use coordinated AI systems to reach market valuations of $1 billion and more. Rocketable, for example, is backed by Y Combinator, True Ventures, and Bloomberg Beta. Andon Labs is backed by Y Combinator and investors including Breakpoint Capital, Phosphor Capital, Superangel and Juniper Ventures.

Limitations Abound

Most so-called autonomous companies still depend heavily on humans for legal accountability, finance, governance, infrastructure, and crisis management. Multi-agent coordination is fragile. AI agents hallucinate (fabricate facts), loop endlessly, misinterpret goals, and sometimes behave unpredictably when given persistent access to tools and memory.

Andon Cafe’s Mona, for instance, sometimes orders far too much bread, and on other days, misses bakeries’ daily deadlines entirely, forcing baristas to remove sandwiches from the menu. Andon Labs’ technical staff attribute this to the AI’s “limited context window”—when old memory of past orders falls out of context, she forgets what she previously ordered.

There’s also the concern over accountability with Agentic AI “bosses”. PocketOS, as an example, was using a Claude-powered agent through Cursor when a routine task spiralled out of control. After encountering a credential issue, the agent attempted to fix the problem autonomously but executed a delete command that wiped out everything including backups—all in just nine seconds. There was no approval step, no verification, and no human intervention. When questioned later, the AI admitted it had ignored its own safety protocols and had simply “guessed” the action, highlighting the urgent need for governance and human control.

By 2028, AI agents will outnumber human sellers by tenfold, but less than 40% sellers will report that AI agents have improved their productivity, according to Gartner. As AI agents become ubiquitous, sales leaders should move beyond simply increasing the number of bots. Instead, leaders should focus on integrating AI into cohesive strategy that prioritise data quality, process automation, and user experience.

Without a disciplined approach, organisations risk overwhelming sellers and missing out on the true benefits of AI, Gartner cautions.

Further, if AI agents eventually run businesses autonomously, the implications could be profound. The white-collar managerial roles may face the earliest disruption, while shareholders could benefit from ultra-efficient, always-on companies with lower labour costs. For society, however, the rise of autonomous enterprises may intensify concerns around inequality, wealth concentration, accountability, and the future role of humans in the economy.

Looking ahead

The evolution of AI-driven enterprises is likely to unfold in three stages: AI-assisted businesses where humans remain in control; AI-operated businesses where agents manage most operations under human oversight; and, eventually, AI-native firms built largely around autonomous systems. The early signs are already visible in algorithmic trading, automated e-commerce, and AI-driven advertising.

“Tracing”, according to IBM, could become critical as AI agents become increasingly autonomous and complex. Unlike traditional software, AI agents break tasks into multiple steps involving model decisions, external tools, and data sources. “Tracing” records every action an agent takes, including which model was used, which tools were called, the inputs and outputs generated, the time taken, and associated cost.

IBM believes this visibility will allow enterprises to investigate failures, audit decision-making, ensure compliance, and optimise performance by identifying slow, expensive, or inefficient workflows. In essence, tracing helps organisations understand exactly how an AI agent reached a particular outcome.

Subscribe now

AI BITS & BYTES

India uses AI for weather forecasting

India has rolled out two advanced weather forecasting products that signal a major AI-driven upgrade to its weather infrastructure.

The first, developed by the India Meteorological Department (IMD), is an AI-enabled “Forecast of Monsoon Advance over Different Parts of the Country”. The second is a pilot “High Spatial Resolution Rainfall Forecast for Uttar Pradesh.”

The AI-powered monsoon forecasting system will issue probabilistic forecasts every Wednesday, predicting monsoon progression up to four weeks in advance. Designed to support farmers across 16 states and more than 3,000 sub-districts, the platform combines AI models, extended-range prediction systems, and statistical techniques to deliver operational forecasts for agricultural planning and preparedness through the Ministry of Agriculture and Farmers’ Welfare’s dissemination network.

The move comes as India rapidly expands its weather-monitoring network. The country, which had just 17 Doppler Weather Radars a decade ago, now operates around 50, with another 50 planned under Mission Mausam. You may read more about this here.

What machines say about future of human jobs

According to a new study, Anthropic’s Claude, OpenAI’s ChatGPT and Google Gemini aren’t aligned on which jobs are the most at risk due to AI.

The study found that these AI models disagreed the most on supervisory roles and jobs which combine cognitive and physical tasks but seemed more aligned when ranking physical jobs. The models also disagreed on the vulnerability of advertising managers and chief executives.

Claude rated accountants as highly vulnerable to AI automation, while Google Gemini assigned the profession a much lower exposure ranking. The models also disagreed on the vulnerability of advertising managers and chief executives.

The researchers also found that the estimates of the impact of AI on employment could change drastically based on the AI model used. Read more.

Ilya Sutskever’s bombshell on ‘OpenAI-Anthropic merger’

Former OpenAI Chief Scientist Ilya Sutskever has testified in court that CEO Sam Altman has a “consistent pattern of lying”.

According to Reuters, Sutskever said he had been considering removing Altman as CEO well before the dramatic boardroom shake-up in November 2023 and had discussed the possibility “for a long time” with former OpenAI CTO Mira Murati. Altman was later reinstated as OpenAI CEO. Read more.

Hope you folks have a great weekend, and your feedback will be much appreciated— just reply to this mail, and I’ll respond.

Edited by Rashmi Sanyal. Produced by Tushar Deep Singh.

In this week’s edition of Mint Tech Talk:

• AI outperforms human doctors in the emergency room

• AI Tool of the Week: Google Gemini's File Generation

• Freshworks cuts jobs, Cognizant shareholder payouts

PREVIOUSLY, ON TECH TALK

Subscribe now

While it is difficult to quantify AI’s impact on the recent assembly elections in West Bengal, Tamil Nadu, Kerala, Assam and Puducherry, its presence was unmistakable. From hologram speeches and AI-generated campaign videos to surveillance systems and even a humanoid robot at a polling booth, AI tools were widely deployed to scale outreach and personalise messaging.

The Tamilaga Vettri Kazhagam (TVK), led by actor-turned-politician C. Joseph Vijay, exemplified this shift. The party used holograms, robots and AI-generated content to expand its reach across Tamil Nadu, targeting first-time voters through virtual rallies and hyperlocal messaging. Tactics included AI phone calls—where a politician’s voice is synthetically recreated to interact with voters in real time—personalised WhatsApp videos, and real-time speech translation. Last September, an AI-generated video of former Tamil Nadu Chief Minister and DMK founder C.N. Annadurai went viral. It recreated his voice to endorse Vijay while criticising the current DMK leadership.

Data-Driven Campaigning

For years, political strategists have used behavioural models, chatbots and multi-lingual tools to refine voter outreach. AI now amplifies these capabilities, enabling campaigns to tailor messaging based on detailed demographic, psychographic and behavioural data, and to adapt communication to local dialects and contexts at scale.

India has been experimenting with such tools for some time now. In 2021, the State Election Commission, Bihar, partnered with Staqu Technologies Pvt. Ltd. to deploy video analytics and optical character recognition to monitor CCTV footage during panchayat elections.

In the 2024 Lok Sabha elections, Gen AI was used to improve voter communication. Tools such as Bhashini enabled Prime Minister Narendra Modi to deliver speeches in multiple Indian languages, expanding his reach across non-Hindi-speaking regions. AI-powered robocalls delivered hyper-personalised messages, while chatbots and voice assistants enhanced engagement.

In Rajasthan, BJP volunteers received personalised WhatsApp videos in which a party leader addressed them by name using voice cloning and lip-syncing technologies.

In Kerala, Congress leader Shashi Tharoor even appeared alongside his AI avatar at a literature festival, offering a novel form of voter interaction.

The Risks are Evident

To be sure, the misuse of data and targeted political messaging is not new.

Cambridge Analytica’s exploitation of Facebook data during the 2016 US Elections remains a cautionary tale. AI intensifies these concerns. The 2024 global election cycle marked a turning point, effectively becoming a testing ground for AI-driven campaigning for many countries that had to deal with deepfakes, AI avatars, and synthetic endorsements from fabricated celebrities.

In India, for instance, political parties reportedly spent ~$50 million on AI-generated content, including deepfakes of deceased leaders and manipulated videos of celebs, misleading voters at scale.

In 2024, doctored videos showed actors Ranveer Singh and Aamir Khan criticising PM Modi while endorsing the Congress. Another AI video falsely depicted Rahul Gandhi being sworn in as prime minister while voting was still underway.

Agentic AI Raises the Stakes

Such systems don’t just generate content, but plan, act, adapt and optimise campaigns in near real time. Imagine a system that tracks constituency-level sentiment, identifies undecided voter clusters, generates tailored video or audio messages, deploys them across channels, and reallocates resources based on response rates—all continuously.

Pieces of this already exist in silos—data pipelines, generative models, automation frameworks, and distribution channels—but what’s missing is full integration and autonomy.

Simply put, instead of today’s dashboards, content teams and ad managers, Agentic AI can tie everything together with minimal human intervention. What used to take days of coordination could happen in hours. Over time, such systems could evolve into always-on “campaign copilots”, continuously nudging strategy based on shifting voter mood.

Because Agentic AI systems can operate autonomously across channels, they can coordinate and influence campaigns far more effectively than human teams. Think AI agents posing as real users—sustaining narratives, responding to counter arguments, and keeping misinformation alive longer than traditional troll farms could.

A January 2026 paper titled How malicious AI swarms can threaten democracy warns that AI is making population-scale manipulation increasingly feasible. LLMs and Agentic AI can push political campaigns with unprecedented scale and precision, dramatically expanding the volume of persuasive content without sacrificing credibility. Generative AI can produce falsehoods that often appear more human than human-written text, while techniques designed to improve reasoning—such as chain-of-thought prompting—can be repurposed to make misinformation more convincing.

The paper highlights a more disruptive risk: coordinated “swarms” of malicious AI agents. By combining LLM reasoning with multi-agent architectures, these systems can act autonomously, infiltrate online communities, and manufacture the illusion of consensus. By mimicking human social behaviour and adapting in real time, such AI swarms could become a potent threat to democratic processes.

That said, even as agentic systems become smarter, they will require rich, high-quality data to be effective. Second, there are platform controls, and social media firms are getting better at detecting coordinated inauthentic behaviour. Third, there are legal guardrails, and authorities like the Election Commission of India are already pushing for disclosure and rapid takedowns.

The ECI maintains that all stakeholders must ensure the responsible and ethical use of digital platforms, in line with the Information Technology Act, 2000, the IT Rules, 2021, and the Model Code of Conduct. Political parties and candidates are also required to clearly label any AI-generated or modified material and disclose its source to maintain transparency and voter trust. During the recently concluded state elections, the ECI took action against over 11,000 social media posts during the campaign—ordering takedowns, issuing clarifications, and filing FIRs—with a significant share reportedly linked to Tamil Nadu and West Bengal.

Finally, if every party deploys similar advanced tools, their effects may partly cancel out. At least, that’s what one can hope for as guardrails try to keep pace with AI advancements.

Subscribe now

AI BITS & BYTES

AI outperforms human doctors

A new Harvard study has found that advanced AI models can outperform doctors in diagnosing patients, including in high-pressure emergency room scenarios.

The researchers pitted OpenAI Inc.’s advanced o1 language model directly against hundreds of physicians across multiple diagnostic touchpoints, finding that the AI consistently outperformed doctors in both diagnosing conditions and planning clinical management. The new research, published in the journal Science, gave 76 clinical cases from the emergency room of Beth Israel Deaconess Medical Center to OpenAI’s o1 model and two expert attending physicians.

On the flip side, Pennsylvania has sued an AI company, saying its chatbots illegally hold themselves out as licensed doctors.

IBM CEO calls for an AI business model

IBM CEO Arvind Krishna believes maximising returns on AI investment requires a fundamental restructuring of business workflows, rather than just adopting new technology. The use of AI within a company typically evolves from the individual contributor to small teams, cross-functional groups and ultimately the entire firm.

As a company progresses from one stage to another, so do the potential returns on AI investment. That’s less about technology alone than it is about updating age-old processes and social dynamics, according to Krishna. Read more.

A digital wardrobe, courtesy Google Photos

Google is bringing AI to the Google Photos app, allowing users to try out clothes right from their gallery app. In a blogpost announcing the new feature, Google said that the ‘Wardrobe’ feature will “soon catalog the clothes you’re wearing in photos and create a digital closet that puts your wardrobe at your fingertips”. The company noted that its new AI-powered feature will automatically catalog the clothes you wear in your pictures to build a fully searchable digital closet. Read more.

Hope you folks have a great weekend, and your feedback will be much appreciated— just reply to this mail, and I’ll respond.

Edited by Rashmi Sanyal. Produced by Tushar Deep Singh.

Smartphones powered by large language models (LLMs) and Agentic AI promise to reshape personal computing with advanced reasoning, contextual awareness, and autonomous decision-making, even as the fate of overhyped devices such as the Rabbit R1 and Humane AI Pin serve as a cautionary note.

In this week’s edition of Mint Tech Talk:

• OpenAI’s rumoured Agentic AI smartphone

• AI Tool of the Week: ChatGPT Images 2.0

• Everyone wants a piece of Anthropic PBC

In May 2025, Sam Altman (who invested in Humane AI) brought on board legendary iPhone designer Jony Ive to build a new class of AI-first devices at OpenAI. And while OpenAI has made no official comment on this topic, OpenAI could unveil its first smartphone by late 2026 or early 2027, with mass production targeted for 2028, according to Ming-Chi Kuo—an analyst with TF International Securities (HK).

The AI smartphone is expected to blend cloud and on-device LLMs, powered by chips optimised for energy efficiency, memory management, and continuous contextual awareness. Partnerships with MediaTek and Qualcomm are reportedly under consideration, as OpenAI seeks end-to-end control over both the hardware and operating system to deliver a “comprehensive AI agent service”, says Kuo.

OpenAI has built a broad infrastructure stack spanning cloud, silicon, and data centres. The company has also outlined plans for a “superapp” that brings together ChatGPT, Codex, browsing, and broader agentic capabilities into a unified, agent-first experience. An “Agentic AI” phone, and a line-up of similar “agentic” devices, closes the loop on the hardware side.

Still, the question remains: Even if OpenAI does launch a ChatGPT phone, how different will it be from today’s Gen AI smartphones?

Devices by Samsung, Google, and Apple already integrate Gen AI deeply into their chips and operating systems, running on-device models via dedicated neural processing units. Smartphones, such as the Google Pixel 10 Pro with its Tensor G5 chip and Samsung Galaxy S26 powered by the Exynos 2600 on a 2 nm processor, highlight how far this integration has come—fast, efficient, and capable of handling meaningful on-device workloads across text, voice, image, and multimodal tasks.

The market is still early but growing. According to Intel Market Research, Gen AI smartphones were valued at about $95 million in 2025 and are projected to reach roughly $414 million by 2034. In 2025 alone, production reached 185 million units at an average price of $558. Demand is being driven by content creation, productivity, and emerging use cases in healthcare, education, and creative fields, while advances in model compression are pushing these capabilities into mid-range devices.

Yet today’s AI remains largely reactive—it waits for prompts. It can summarise, generate, and assist, but it does not autonomously manage tasks like travel-planning or subscription optimisation end-to-end. A true Agentic AI device, by contrast, would embed intelligence into the operating system itself, shifting interaction from “open an app” to “state a goal”.

Potential buyers?

OpenAI potentially has a captive audience. Currently, it has more than 900 million weekly active users, and that base is expected to approach one billion in the near term. Likely AI-native professionals already using tools like ChatGPT, Gen-Z driving cultural adoption, and enterprises seeking workflow automation. A phone purpose-built for agentic enterprise workflows—one that can autonomously draft, send, and follow up on communications—is an automation tool, not merely a handset.

Cost and trust

Gen AI adds an estimated $120–$200 to device manufacturing, even as privacy concerns remain high, according to Intel Market Research. Layer in EMIs, data plans, and AI subscriptions, and pricing becomes tricky. That is unless, as Google and Apple suggest, AI is bundled and largely hidden from the user. IPO-bound OpenAI, on its part, has not disclosed any details.

But where’s the money?

Meanwhile, The Wall Street Journal reported that OpenAI missed its own targets for new users and revenue, raising concern among company leaders about whether it will be able to support its massive spending on data centres. Chief Financial Officer Sarah Friar has said that she is worried that OpenAI may not be able to pay for future computing contracts if revenue doesn’t grow fast enough.

OpenAI has also ended its exclusivity arrangement with Microsoft Corp., meaning the Azure parent will no longer have sole access to OpenAI’s models and products. The move opens the door for OpenAI to work with rival cloud platforms such as Google Cloud and Amazon Web Services.

Announced on 27 April, the revised agreement also removes Microsoft’s obligation to pay a revenue share on OpenAI products it resells via its cloud.

“The greater predictability in the amended agreement strengthens our joint ability to build and operate AI platforms at scale, while providing both companies the flexibility to pursue new opportunities,” the two companies said in a joint statement.

Subscribe now

AI BITS & BYTES

Everyone wants a piece of Anthropic

Google will invest $10 billion in Anthropic PBC, with another $30 billion potentially to follow, strengthening the relationship between two firms that are at once partners and rivals in the race to build AI.

Anthropic has ramped up its fundraising amid the breakout success of Claude Code, an AI agent that speeds the process of writing computer software. The startup said earlier this week that it nabbed another $5 billion from Amazon, also at a valuation of $350 billion, with the option to inject another $20 billion over time.

Anthropic raised $30 billion in February and investors have since sought to back the firm at a valuation of $800 billion or more. Read more.

AI exposure highest in skilled roles not low-wage jobs

According to Anthropic’s Labor Market Impacts Report, professionals currently most exposed to AI are those with higher education, significant experience and above-average earnings.

The findings in this report turn conventional thinking upside down. There’s a clear tension here: the qualifications and expertise that once helped secure a stable career may now be the very factors placing workers at greater risk of disruption.

Amazon’s 6-point AI playbook for engineers

Amazon.com Inc. is formalising how it builds with AI as part of a broader push to make AI central to its engineering culture, Business Insider reports. The e-commerce firm’s massive retail division, known internally as “Stores”, has formalised its approach into a set of six “AI-native engineering tenets” designed to guide how teams should approach AI development across the organisation.

Meta to unwind $2.5-billion Manus acquisition

Meta Platforms Inc. is planning to unwind its acquisition of AI startup Manus after the Chinese government banned the transaction on national security grounds, according to a new report by The Wall Street Journal, citing sources familiar with the matter.

Any attempt to reverse the deal is expected to be complex, requiring the company to disentangle operations, data, and technology that have already been combined.

Meanwhile, a Bloomberg report noted that Manus staff have been integrated into Meta Platforms, funds have been dispersed, and the startup’s leadership has joined the company’s AI division.

Elon Musk tells his side of the OpenAI story

Elon Musk took the stand for the second day Wednesday in the landmark trial that pits the world’s richest person against Sam Altman, a fellow OpenAI co-founder he accuses of betraying promises to keep the company as a non-profit dedicated to humanity’s benefit.

Musk, who invested about $38 million in OpenAI from December 2015 through May 2017, gave his account of OpenAI’s early years, recounting how he lost confidence that Altman would keep it a non-profit. Questioned by his lawyer Steven Molo, Musk said by late 2022 he was concerned Altman was trying to “steal the charity.”

Hope you folks have a great weekend, and your feedback will be much appreciated— just reply to this mail, and I’ll respond.

Edited by Rashmi Sanyal. Produced by Tushar Deep Singh.

In this week’s edition of Mint Tech Talk:

• Hiring slows but data shows AI isn’t killing jobs

• Humanoid beats human record at half-marathon

• AI Tool of the Week: Skills in Chrome

• ChatGPT Images 2.0 ups the stakes

Meta Platforms Inc.’s Mark Zuckerberg, for instance, is working closely with software development teams after two decades. He has moved his desk into Meta’s AI Lab, sitting alongside Alexander Wang and Mat Friedman—co-leads of Superintelligence Labs—and has been actively writing and reviewing code, according to a report by Business Insider. Meta is even building a photorealistic, AI-powered 3D avatar of Zuckerberg to engage with employees in his stead, Financial Times reports.

Google co-founder Sergey Brin is making a similar return. The Information reports that the Gemini maker is assembling an elite “coding strike team” to close the gap with Anthropic, with Brin directly involved. The ambition is to push towards an “AI takeoff” in which systems increasingly improve and code themselves.

Closer home, in January 2025, co-founder Sridhar Vembu stepped down as CEO of Zoho to become chief scientist, focusing once again on writing and reviewing code.

Set against this, Anthropic co-founder and CEO Dario Amodei has argued that coding itself may eventually disappear, replaced by systems that turn human intent directly into working software.

Anthropic is already building toward that future with tools like Claude Code, Mythos (restricted distribution due to security reasons even as OpenAI Inc.’s GPT-5.4-Cyber, released this month, is a model finetuned for additional cyber capabilities and with fewer capability restrictions for those who are part of its Trusted Access for Cyber programme), and Claude Design.

After triggering a selloff in stocks of SaaS companies earlier this year with Claude Code, Anthropic’s new AI tool Claude Design is threatening software design firms like Adobe and Figma. Anthropic says Claude Design can create polished visual assets ranging from interactive prototypes to pitch decks—all with a simple text prompt.

Given these developments, what looks like a contradiction is, in fact, a shift. AI is reshaping how software gets built. Instead of writing detailed instructions, developers increasingly describe outcomes and let machines generate the rest. That makes AI the new interface between humans and computers—a role once held by operating systems and cloud-computing platforms. Whoever controls this interface, in effect, shapes how software is created.

This is precisely why founders are getting closer to software again—not just to write more code but to understand how this new layer behaves.

Consider Meta’s internal targets. According to documents cited by Business Insider, the company expects that by mid-2026, 65% of engineers in its core product groups will generate more than 75% of their code using AI. In a world where small changes in prompts can produce widely different outcomes, rolling up one’s sleeves is not a throwback. It is a strategy.

At the same time, AI is compressing software work. Tasks that once required entire teams can now be handled by a single developer equipped with the right tools. This is not full automation, but a reduction in the number of people needed between idea and execution. Cost too is becoming a constraint: AI-generated code is not free. Tokens carry a price. Every prompt and iteration consumes compute, raising a basic question of efficiency: When is it cheaper to use machines, and when do humans remain more effective?

The answer will shape how quickly AI transforms traditional development. For now, the technology remains imperfect. AI can generate code quickly, but it is still prone to errors, edge cases, and unpredictable behaviour. That makes human oversight more important, not less. The work is shifting from writing code to guiding and validating it.

So what changes? Routine coding tasks will decline. Large engineering teams may shrink. But higher-order skills—system design, judgment, and the ability to translate ideas into precise, executable instructions—will become more valuable. Coding does not disappear; it moves up the stack. The founders returning to it understand this. The future may need fewer programmers, but it will demand more leaders who can think like one.

Subscribe now

Hiring slows but data shows AI isn’t killing jobs

The bad news for white-collar workers continues. As per layoff.fyi, a little over 81,000 employees have been laid off by 97 tech firms since the beginning of this year.

LinkedIn too has confirmed a 20% decline in hiring since 2022. But it does not blame AI for the ongoing slowdown, as per a TechCrunch report, citing an interview at the Semafor World Economy Summit. LinkedIn’s Chief Global Affairs and Legal Officer Blake Lawit instead said the broader decline in hiring is more closely tied to rising interest rates.

“At LinkedIn…we have an economic graph which is over a billion members. We’ve got companies, jobs, skills. It’s really an amazing real-time view of what’s happening in the labour market. And we’ve looked—because everyone wants to know the answer to this question: Is AI impacting jobs right now? We’ve looked and, honestly, we haven’t seen it,” Lawit said. He argued that if AI was actively replacing workers, the job losses would be highly visible in specific, vulnerable sectors such as customer support, administrative roles, or marketing, which was not the case.

Even Stanford computer science graduates “are struggling to find entry-level jobs” in “a dramatic reversal from three years ago”. Yet, it’s hard to find data that provides evidence linking AI adoption to reduced job postings, according to an article in The Stanford Review.

Meanwhile, Yann LeCun, former AI chief scientist at Meta and founder of AMI Labs, has criticised Anthropic’s Amodei for his remarks that AI will reduce tech jobs by 50%. Lecun noted that AI lab CEOs—including himself, OpenAI CEO Sam Altman, Yoshua Bengio, and “Godfather of AI” Geoffrey Hinton—aren’t necessarily the right authorities to forecast AI impact on jobs and the broader economy. Instead, Lecun noted that economists can be a better judge of how AI will affect the job market.

“Dario is wrong. He knows absolutely nothing about the effects of technological revolutions on the labour market. Don’t listen to him, Sam, Yoshua, Geoff, or me on this topic. Listen to economists who have spent their careers studying this.”

By 2030, Gartner says, CIOs will expect all IT work to involve AI, with 75% handled by humans augmented by AI and the remaining 25% performed autonomously. The research advisory maintains that AI is less about job loss and more about workforce transformation, and expects AI’s overall impact on global employment to remain neutral through 2026 and turn net-positive by 2027 as new roles outpace those displaced.

AI BITS & BYTES

Humanoid beats human record at Beijing half-marathon

A humanoid robot developed by Honor has captured global attention after winning a half-marathon race for robots in Beijing, posting a time that surpasses the human world record for the distance. The event, held in the Beijing Economic-Technological Development Area (E-Town), highlights China’s accelerating progress in artificial intelligence and robotics.

The winning robot completed the 21-km course in 50 minutes and 26 seconds. This time is notably faster than the current human half-marathon world record held by Jacob Kiplimo, who ran the distance in approximately 57 minutes at a Lisbon road race in March 2026.

A separate Honor robot—operated via remote control—crossed the finish line even faster at 48 minutes and 19 seconds. However, under the competition’s weighted scoring system, which prioritised autonomy, the fully autonomous humanoid robot was awarded first place.

ChatGPT Images 2.0 ups the stakes

The internet is buzzing over OpenAI’s rollout of ChatGPT Images 2.0, which it frames as the start of “a new era of image generation”. The update builds on the December launch of ChatGPT Images positioned as a response to Google Gemini’s Nano Banana.

ChatGPT now “thinks” before it creates—planning outputs, pulling in references, and checking for errors before delivering results. The upgrade also brings practical gains: up to 2K resolution, batch generation of eight images, flexible aspect ratios from 3:1 to 1:3, and improved multilingual text rendering. It is not perfect, though.

I prompted it to “Please draw a picture of a Royal Enfield cruiser bike with essential accessories.” What I got was a decent picture but the fuel tank, as you see, belongs more to the Classic Model rather than the Meteor 350.

As ChatGPT Images 2.0 works to improve with time, the move intensifies competition in a crowded field that includes Midjourney, Stable Diffusion, Adobe Firefly and Google Gemini.

GPT 5.5 ‘Spud’ looms, GPT-6 still a meme

Reports and speculation suggest ChatGPT 5.5 (codenamed “Spud”) will be a modest upgrade within the GPT-5 family, aimed at bridging the gap between GPT-5.4 and GPT-6, with a possible release by April end. OpenAI, however, has made no official announcement.

As for GPT-6, the signal is even noisier. In October 2025, CEO Sam Altman leaned into the viral “6-7” meme, joking on X that GPT-6 might be rebranded “GPT-6-7,” after Dictionary.com crowned the absurd phrase its Word of the Year. The comment was tongue-in-cheek but underscored how Big Tech firms make gossip work for them.

Hope you folks have a great weekend, and your feedback will be much appreciated— just reply to this mail, and I’ll respond.

Edited by Rashmi Sanyal. Produced by Tushar Deep Singh.