Good Morning,

I envision a world where LLMs are complemented by many types of AI, Quantum computing and a convergence of new kinds of computing architectures. For all the book smarts of LLMs, they currently have little sense for how the real world works and there’s a growing list of “world model” AI startups trying to fill the gap. People won’t trust a one dimensional AI, more on this at the end.

While Generative AI models have their uses, I believe the next decade will fill in some of the physical AI gaps, and allow for things that were previously not possible. Yann LeCun, Fei-Fei Li and others are working towards this end. The likes of Google, Meta, Tencent, Nvidia and others are of course also developing world models.

Robotic brain startups like Physical Intelligence (π) are building general purpose models for physical AI that are basically foundational models for the physical world to help robots learn among other things. Language isn’t enough and models trained on text have inherent limitations. The token based way of predicting the world is just one perspective. A world-model based AI could have advantages over LLMs that understand casualty, and maintain mental models that are more human-like and grounded. This could enable more efficient decision making AIs, to actually make the world a better place inspired by cognitive neuroscience and grounded in empirical experience.

Yann LeCun, , Packy McCormick and many others make a lot of good points about the inherent limitations of LLMs and the merits of World Models, and I think this is important to point out even as Silicon Valley fuels capex that isn’t rational compared to what the future AI will actually be like. Of course that is just my opinion.

My guest contributor today is of Weighty Thoughts Newsletter, a VC on AI and DeepTech. 📚 His book What you Need to know about AI, recently won this year’s American Legacy Award Best Nonfiction . It was also a finalist in the business technology category. Congratulations to James.

Weighty Thoughts ∘˙○˚.•

Weighty Thoughts

VC on AI, deep tech, startups. Former Bridgewater, Google[x], startup founder. Read by top engineers, fund managers, and policymakers.

By James Wang

My Weights Thoughts of the day ⚗️

A little Quantum tangent on our World Models exploration today:

• Nvidia recently announced Nvidia Ising that the company says might be the world’s first open AI models to accelerate the path to useful Quantum computing. It’s exciting because it’s being piloted by the likes of Academia Sinica, Fermi National Accelerator Laboratory, Harvard John A. Paulson School of Engineering and Applied Sciences, Infleqtion (that is now a public company under the ticker ) , IQM Quantum Computers, Lawrence Berkeley National Laboratory’s Advanced Quantum Testbed and the U.K. National Physical Laboratory (NPL).

Read more

Read: Where's your Ed At

I’m paying special attention to the details around AI at the World Economic Forum event in Davos. But the most interesting thing so far in AI in 2026 is that the internet is increasingly questioning the business model of companies like OpenAI. Recently OpenAI Chief Financial Officer Sarah Friar said in a blog post on Sunday that the company's annualized revenue has surpassed $20 billion in 2025, up from $6 billion in 2024 with growth closely tracking an expansion in computing capacity. Anthropic’s meanwhile has hit a run rate of already $9 Billion and will increase faster in 2026. Anthropic is a full six years younger than OpenAI.

Generative AI skeptics like Gary Marcus are going even more viral in 2026 while Ed Zitron has built a growing reputation built on OpenAI skepticism and AI Bubble explorations. But against a tidal wave of AI boosterism, the tweet length shots of Marcus or the Bible length passages of Zitron’s own anxious persuasion aren’t necessarily the best evidence of today’s AI reality.

The FT reported this week that speaking at the World Economic Forum at Davos, Switzerland on Tuesday, that Microsoft CEO Satya Nadella pontificated about what would constitute such a speculative bubble, and said that the long-term success of AI tech hinges on it being used across a broad range of industries — as well as seeing an uptick in adoption in the developing world where it’s not as popular. The data on this adoption is not actually very good as we will dissect further major CEO survey that recently dropped.

While OpenAI will push out its ChatGPT enabled earbuds later in 2026, its comes on a rather grim backdrop of adoption. It’s not clear if ChatGPT’s so-called 800 million “weekly” active users will care, compared to the smart glasses from Google, Apple and Meta. But let’s take a step back into that adoption problem.

Adoption rates of Generative AI in real companies and OpenAI’s own losses aren’t adding up into a very coherent path to profitability. The financial numbers are bleaker than you can possibly imagine.

OpenAI's computing capacity rose to 1.9 gigawatts (GW) in 2025 from 0.6 GW in 2024 that came with a price tag that it’s not clear when OpenAI, if ever, will able be able to afford as Oracle takes on incredible risks of things like long-term Datacenter leases.

The Microsoft-OpenAI Files paint a bleak picture of Microsoft’s role in pushing OpenAI into a lack-lustre product company that cannot even deliver good products. New court filings offer an inside look at one of the most consequential relationships in tech and how things can go terribly astray. On one hand we have OpenAI CFO Friar saying OpenAI will focus on practical applications of ChatGPT, while in the same week they push more into ChatGPT “adult mode”. It’s a stark and comical contrast.

“For this not to be a bubble by definition, it requires that the benefits of this are much more evenly spread,” Nadella said. The “tell-tale sign of if it’s a bubble,” he added, would be if only tech companies were benefitting from the rise of AI.

Of course in 2026, that’s exactly what we are seeing. Outside of use cases like coding and at law firms. Evercore ISI's Mark Mahaney wrote advertising, Ads in ChatGPT, could be a $25 billion business for OpenAI by 2030. The problem is of course that’s not nearly enough compared to how gigantic OpenAI’s losses are going to widen into as it descends into the valley of death in the 2026 to 2029 period.

Subscribe now

Read more

The opening act of 2026 has felt fairly intense in the AI news cycle. Three days ago Google introduced Gemini’s Personal Intelligence. OpenAI is finally expanding into Ads in ChatGPT. Five days ago Anthropic pushed out Cowork. It led to a lot of chatter on Twitter/X.

Cowork was obviously inspired by success of Claude Code and its flagship Opus 4.5 model, that was one of the biggest improvements in AI in all of 2025. We’ve been uncovering the pace of change in Generative AI with three featured articles: In case you are a bit slow to get back into work this year, here’s what you missed:

1. 2025 Recap: The Year the Old Rules Broke

2. Key Milestones of China in AI of 2025

3. AI Trends: 2025 Lookback and 2026 Outlook

I asked to help us break us down Anthropic’s recent product success and their transition from Claude Code to trying to make the desktop version a bit more accessible with Claude Cowork. His analysis of Claude Code is fairly deep with tons of details.

Gearing up for historic AI IPOs ✨

Two days ago Anthropic published their 4th edition of their Economic Index. Cowork was built using Anthropic’s own tools so you have to wonder if this is a side note in the ramp up to a kind of echo of recursive self-improving AI in product? It’s also branding right, Anthropic’s chief scientist has said this pseudo AI autonomy could spark a beneficial ‘intelligence explosion’. Jared Kaplan said that while efforts to align the rapidly advancing technology to human interests had to date been successful, freeing it to recursively self-improve “is in some ways the ultimate risk, because it’s kind of like letting AI kind of go”. Will Cowork be remembered as a first of sorts?

It’s about that time where you need to follow someone called “Boris” who works at Anthropic. Google’s own Personal Intelligence with Gemini is also an upgrade.

Did Gemini just make its “Her” Move? 👩🏻

Google has an absurd ecosystem advantage so what happend if you added “memory” to that? Gemini Personal Intelligence is a new capability layer from Google that allows Gemini to act as a unified reasoning engine across your entire Google ecosystem. People are racing to that “HER” (movie) style capability and many ChatGPT users are reporting switching to Gemini in recent months.

Google with Personal Intelligence can reason across: ✨

→ Your Gmail threads ✉︎
→ Your Photos library 📷
→ Your Search history 🔍
→ Your YouTube watch patterns 🎬

That really could lead to a more personalized AI over time. This is all planned ofcourse to coincide with Google’s AI-powered smart glasses that will integrate Gemini and Personal Intelligence. This is timed to compete with Apple’s AR glasses and OpenAI’s AI wearables. So in 2026 we’ll see not just BigAI IPOs but exciting wearables related to these chatbots urging us to take them with us wherever we go.

Boris Cherny has a steady stream of Claude Cowork tidbits, like improvements they are adding daily.

All in all it really does look like AI might be a bit more actionable and useful in 2026. We already covered how the agent protocols enable this to happen. Google, Anthropic and Qwen (Alibaba Cloud) appear to be at the forefront of making AI more useful.

Anthropic’s Claude Code was the AI product of the Year of 2025 for me. Ilia’s article really goes into some depth about why this is meaningful for the future of AI. A lot has been said on this platform as well around Anthropic’s leadership in AI and about Claude Cowork.

What’s all this about this guy named Claude? 🚨

• Official Claude Prompting: Everything You Need To Know

• The 5 Claude Skills Every Digital Product Creator Needs

• Your Guide to Claude Cowork

• Claude Coworks

• Anthropic Shipped Cowork in 10 Days Using Its Own AI

• Claude Code Coded Claude Cowork

Clearly Claude Code and the Cowork launch was in the air. 💥 (though I have my eyes more on DeepSeek to be honest).

Lydia is another X account to follow regarding Claude Code and Cowork. What all of these articles don’t say exactly is how impactful Opus 4.5 has been in tandem with the Claude Code launch. This is leading Anthropic to develop their product more rapidly than their competition as well as having the best in class AI coding models. This competitive advantage in business adoption might turn out to be profound.

One of the best articles to understanding how Claude Code was built is this one.

Claude Code: Deep Dive

The Accidental $1 Billion Product (And Why It’ll Keep Growing)

Ilia writes Prosper, a newsletter about practical AI workflows and systems thinking. He cover how professionals can use AI tools effectively.

Prosper

The AI newsletter teaching you practical AI workflows, systems, and strategies to think better with AI – not just work faster.

By Ilia Karelin

The Side Project That Became a Billion-Dollar Business

In September 2024, Boris Cherny joined Anthropic with a simple mandate: prototype new products.

Cherny had an unconventional background for a tech leader. He studied economics, dropped out to start companies at 18, worked at a hedge fund, and spent a decade climbing from IC4 to Principal Engineer at Meta - all without a computer science degree. What was surprising, is that Boris didn’t have a computer degree. At Anthropic, his first creation was almost comically mundane: a command-line tool that used Claude to identify what music was playing during work sessions.

Then a colleague, Cat Wu, suggested adding filesystem access.

“Suddenly, this agent was really interesting,” Cherny recalled. Claude could now autonomously explore codebases, read files, and follow import chains. What started as a musical curiosity had become something far more powerful.

Within five days of internal release in November 2024, 50% of Anthropic’s engineering team was using it daily. By July 2025, the tool processed 195 million lines of code weekly across 115,000 developers. In November 2025 - just six months after public launch - Claude Code hit $1 billion in annualized revenue.

The accidental product had become Anthropic’s fastest-growing business.

Editor’s Note: Anthropic’s ARR is highly likely to surge in 2026 in dramatic fashion. As both Anthropic and OpenAI are set to go IPO in the months ahead this is all getting rather serious for the future winners of AI. Claude Code is central to to that and what comes next:

What Makes Claude Code Different

Read more

Hey Everyone,

Welcome to another article from AI Supremacy, a Newsletter about AI at the intersection of tech, business, society and the future. Wel…

Read more

Hey All,

What I’m reading this week:

• Fed’s Paper on Generative AI at the Crossroads

• Microsoft’s report on The New Future of Work (more on the actual 2025 paper below).

• Menlo’s Mid Year Market Update

AI News this Week

• Anthropic released Claude Opus 4.1, that’s even better at coding.

• OpenAI released its open-weight models gpt-oss

• Meta is doubling down on AI, with Mark Zuckerberg declaring that “superintelligence” is now in sight as the company’s systems begin to improve themselves.

• Google DeepMind released its world model Genie 3. World models of 2025 are going to accelerate robotics.

• Google releases their coding agent Jules.

• Manus A.I. launches Wide Research

• Google launches Deep Think in Gemini app

• NotebookLM announced video overviews.

• Google DeepMind launches a Unified AI Model for Planetary Mapping called AlphaEarth (Astronomy).

Audio rundown on AI news items July-August, 2025

5 minutes, 50 seconds.

Big Questions about AI Linger 🧠

What happens when you begin to question everything around AI? While GPT-5 is delayed yet again due to capacity issues among other things - some of the more technologically optimistic narratives are starting to break down, in contrast to significant investments:

• The datacenter roll out too has its risks, like OpenAI’s valuation growing too fast.

• Eric Flaningam updated the market cap update of BigTech: Capex Keeps Rising!

• Among the four main BigTech players, CapEx ~ has reached $88B this quarter! This signals 2026 will push demand for compute and AI chips even higher.

Capex vs. Reality 🌊

• AI is supposed to benefit what young people can do, but what if it’s young people who get caught in a labor crunch?

• There are signs of a hiring pullback in the technology sector, hitting younger employees there the hardest, according to Joseph Briggs, senior global economist of Goldman’s research division.

• Unemployment among Gen Z workers in the tech sector has risen faster than the sector’s overall unemployment rate and that of young workers in general, according to Joseph Briggs, senior global economist for Goldman Sachs’ research division. Why do I get the feeling BigTech doesn’t care about that?

• Unemployment rates among tech workers between 20 and 30 years old jumped by 3 percentage points since the start of this year, Briggs said. That’s a lot, especially when you consider the capital required to build all of these datacenters, their environmental costs and suddenness of the investment

• The Financial Times notes that these debt-funded infrastructure investments were based on fantastical demand projections for a product that quickly became commoditised.

BigTech Capex is being rewarded by shareholders in a race for AI Supremacy

• Meanwhile OpenAI is itself worth $500 billion before it even makes $20 Billion, many years from even being profitable. What could go wrong? Historically this is a very high number. While OpenAI’s revenue is growing fast, most of this is based on the most fickle revenue of all - subscriptions from consumers.

• Other major BigTech companies hit $500 billion valuation when they were far more mature businesses.

  

• OpenAI’s historic revenue growth speed is pushing the world to build more and more datacenters. But it’s probably the wrong signal to follow.

Are Data Centers the New Offices?

We are pouring money into datacenters and to meet demands for compute that can be artificially prolonged indefinitely to grow the Cloud even faster, but is that an ideal scenario? Does it benefit civilization?

“We're getting extremely close to the day when America spends more money to build data centers than office buildings Only 5 years ago, office construction spending was 7x data center spending!” - Joey Politano on X.

Early Adopters vs. the Rest of Humanity

A recent Gallup report showed that it’s mainly white-collar workers increasing AI-use with 27% frequently using AI at work, up 12% since 2024.

• Most of AI use is in fact concentrated in three fields. Industries with the most prolific AI users were tech (50%), professional services (34%), and finance (32%.)

💡 What if? 🤔

• Generative AI doesn’t impact GDP?

• Generative AI doesn’t really boost productivity in a meaningful way as an economic driver?

• BigTech and Cloud hyperscaler capital expenditures in AI Infrastructure, AI talent, models and products never deliver full ROI?

• What if Gen AI isn’t a General Purpose Technology - like we have been led to believe?

What if the gulf between the reality and the hype is so manipulated, even Washington has been deceived. What will be the consequences of all of this? Now I’m not suggesting AI is entirely a money trap, but this does not sound like new technological layer that has much due diligence, risk management or even positive impact on society outside of making the most powerful companies grow faster.

Meanwhile, frequent use of AI by production and frontline workers was flat for two years, from 11% in 2023 to 9% in 2025, per Gallup.

There are two sides to this story, odd that the other one isn’t being told.

Even as Generative AI proliferates as a disruptive technology, what if its economic impact just aren’t as significant as we are being led to believe. I haven’t read much analysis or many projections of what happens if AI doesn’t reach the enormous potential we have set for it.

General Purpose Technologies are characterized by pervasiveness, inherent potential for technical improvements, and ‘innovational complementarities’, giving rise to increasing returns-to-scale. Are we starting to see that with Generative AI? What if we are over-hyping a technology that initially benefits the few, rather than the many? I cannot shake some of these big questions around AI. Not everything is adding up from the vision down to the reality. And if we are not seeing the promise materialize, how long do we wait expecting to see the projected results? And if it doesn’t materialize, how long will we be stuck pretending. The Microsoft paper (more on this later) repeated themes we’ve read before:

Top occupations with lowest exposure to AI

U.S. Labor force disruptions - White vs. Blue Collar

I haven’t read many stories about the blue-collar boom, a reaction to uncertainty. blue-collar roles have evolved in the past few decades and have become more attractive career paths for young people, and with the low risk of AI exposure, blue-collar jobs will only become more popular. In the U.S. as youth unemployment rises, immigration slows (read paper) and AI is being used to target citizens like for deportations, what sort of a world is AI really leading us to?

• How can we expect AI to moderate the economic impacts of slower immigration? The total US population is growing at an annualized rate of 0.5 percent, down from 1 percent in late 2023.

• Clearly many of our expectations around the U.S. economy and the impact of AI aren’t in sync with the reality on the ground.

• Negative economic and jobs data is beginning to be suppressed by the White House, as the backwards looking data really isn’t great in terms of having a strong economy or abundant opportunities for decent jobs. AI isn’t going to save us from ourselves. Productivity boosts from AI are virtually non-existent for a multitude of real world jobs. Large segments of the population have no use for Generative AI or their products in their jobs. While they are being exaggerated in the most cited use cases.

If AI was a general purpose technology, I’d expect spending to approach 19th century Railway (road) like levels. To have measurable and significant impacts on GDP. Three years later that’s not what we are seeing.

What we are witnessing in 2025 is mostly just OpenAI adoption of ChatGPT becoming more popular (globally) and AI coding’s impact with Claude code. In the United States the AI Boom’s risks to the economy are barely even talked about. Skepticism is shrugged off as ignorance to dominant group-think of Techno-optimism (literally the manifesto of VCs), that is now the politically correct dominant ideology to have. Analysts on Wall Street are also starting to spin some fairly speculative projections:

Will BigTech Capex on AI Infrastructure have a Material Impact on U.S. Gross Domestic Product?

Read more

Share

Thinking Machines Lab officially closed a $2B seed round at a $12B valuation – the largest seed round in VC history and 4x bigger than any previous seed funding. The era of Gen AI calls for moonshots, and Thinky is unusual in many regards.

In a world where women struggle in Technology, to be CEOs or even to get funding from mostly male VCs, Mira Murati is the exception. She breaks all the rules, and from Albania to Silicon Valley, we are witnessing a special moment in history. 🏹

The scope of the project is also related to Open-source LLMs, a vast wasteland in the West where the likes of Mistral, Together AI, Ai2 and a few others huddle around the campfires of open-weight innovation with Hugging Face, hoping for better days.

Well, it’s official now, Thinking Machines Lab has raised $2 billion in fresh capital — making it the largest seed round in venture capital history for a company that's been around for just six months and has yet to ship a single product. But don’t let that fool you!

Thinking Machines Lab is less than a year old and has yet to reveal what it’s working on.

I’m so passionate about AI startups and their potential, but even this story is hard to fathom at the peak AI bubble in the United States, where even the Pentagon wants in on the action. But when you dig a bit deeper it actually starts to make a bit more sense.

• Investors include a16z, NVIDIA, Accel, Cisco, AMD, Jane Street

• Murati has said that their first multimodal, open-source-friendly product lands “within months”

Here is a startup that the two U.S. leaders of AI chips are both involved in, Nvidia and AMD, along with a16z making a classic moonshot bet.

The quality of the AI talent at Thinking Machines Lab is likely to exceed that of Google Deepmind or OpenAI, only approached by Anthropic. Because we are living in the OpenAI mafia era where splinter projects become more attractive both for doing pioneering work and for the financial incentives.

Thinking Machines Lab represents yet another “pivot” in the failed Mission of OpenAI. While Google, xAI, Meta Superintelligence Lab and even Anthropic mirror the tactics and behavior of OpenAI, here we have yet another fresh start of OpenAI vets thinking and planning for a re-do. If OpenAI was founded out of fear and envy of Google DeepMind, a lot has changed in just a few years. Thinking Machines know that they can do better, that LLMs weren’t supposed just to be people-pleasing token gulping sycophant-heavy machine economy chatbots.

The Product is a Mystery

Murati peeled back the curtain on the company’s first product a bit in a post on X on July 15th, 2025, claiming that the startup plans to unveil its work in the “next couple months,” and it will include a “significant open source offering.”

At a time when China has overtaken the U.S. in open-weight models, it’s a bizarre period of the evolution of LLMs where new frontier models land nearly every month or at most, every six weeks.

Amid OpenAI’s controversial past, Mira Murati is a recognisable face - Murati rocketed into the spotlight in 2023 when she was named interim CEO of OpenAI after Sam Altman was briefly ousted by the company’s board. It was a complicated time where ChatGPT went on to become synonymous with AI, and only Anthropic, itself a splinter startup of OpenAI’s original team, seemed to be able to follow. So now we have a third player in the game.

This is Talent

The team is comprised so far of about 30 researchers and engineers, with roughly two-thirds being former OpenAI employees, alongside talent from Meta AI, Mistral AI, Character.AI, and Google DeepMind. When a mission goes south like it did at OpenAI, what is the phoenix that rises from the ashes of manipulation and mismanagement (not to mention stringent NDAs, and a murder framed as a suicide?). Some of these names might sound familiar to you:

• Mira Murati, CEO and founder, formerly OpenAI's Chief Technology Officer, instrumental in developing ChatGPT, DALL-E, and voice models.

• John Schulman, Chief Scientist, an OpenAI co-founder who helped build ChatGPT and previously worked at Anthropic on AI alignment.

• Barret Zoph, Chief Technology Officer, former OpenAI Vice President of Research, who left OpenAI alongside Murati.

• Jonathan Lachman, former head of special projects at OpenAI, now a key member of the founding team.

• Lilian Weng, a co-founder with a background in AI safety and robotics at OpenAI.

• Alexander Kirillov, former multimodal research head at OpenAI, now part of the leadership team.

• Andrew Tulloch, with expertise in pretraining and reasoning from his time at OpenAI.

• Luke Metz, a research engineer with experience in post-training at OpenAI.

• Devendra Chaplot, a founding team member, previously involved in AI research.

• Myle Ott, another founding team member with a strong research background.

In some ways Thinking Machines Lab is more OpenAI than the talent who work at OpenAI are today - and might need to take that original mission and fulfil it with a greater promise. These are some of the people that witnessed OpenAI become almost unrecognizable in just a few short years of product and commercial pressure.

If I had to take key values or framing of what this startup is about from their website, I’d gather:

• Science

• Open-Source

• Democratization of AI

• Explainability of frontier AI systems

• Customizable models at scale

You will notice the mission statement is less ideology and more pragmatism of action. Less Sam Altman-Ilya Sutskever cult-speak and more holistic alignment.

A $12 Billion valuation for a startup that really doesn’t formally exist yet with a product, is totally unheard of and the biggest moonshot we’ve yet to see in the AI space. I’d argue it’s more important also than Meta Superintelligence Lab trying to poach talent from OpenAI, Apple and xAI in some reckless act of Tycoon envy otherwise known as a day in the life of Mark Zuckerberg.

Thinking Machines isn’t a long shot, it’s the highest average concentration of AI talent we’ve ever seen in a small deal, higher than DeepMind in its early days. Led by a woman, and that’s fundamentally important when building and aligning an early team. These are OpenAI vets that have seen the good and bad of what power and influence can do to a rising AI startup. Who have witnessed Sam Altman in his paranoia and bizarre attempts at indoctrinations in a mission he doesn’t even truly believe in. Mira Murati isn’t Sam Altman, and it matters.

Thinking Machines Lab - Ghost in the Shell

This is a research lab unlike any other on the planet, perhaps even a radical departure from commercial norms like DeepSeek brought us six months ago.

These are also seasoned Machine learning researchers who know the responsibility and the weight that building LLMs can have on the world, who understand their impact like few other human beings on the planet.

A Little bit more of “Thinky” in our Life

The acronym or pet name on X for the startup is quickly morphing into “Thinky”. It appears to have a more action biased plan than SSI, the other mega “AGI startup” in the OpenAI Mafia leagues.

Mira Murati has been working diligently, she left OpenAI in September of 2024 and launched Thinking Machines in February, 2025 though she has not shared many details about the startup publicly.

Interesting corporate backers like Servicenow and Cisco are surprising. As was the fact the valuation is at $12 Billion and not $10 Billion that was earlier reported. This is a project many VCs wanted to get into even at an astronomical valuation. Consider the uniqueness of the merits of the team, the average seed round size for AI startups in 2024 was approximately $3.5 million, according to industry analyses. This team managed to get $2 Billion?

These are clearly “big” thinking machines that the startup are working on. Amid OpenAI’s delay of their own open-weight model, it gives more weight to the work that Thinking Machines are doing behind the scenes. With Meta going closed-model now, it creates a huge void in the ecosystem as Chinese models take the crown in the open-weight LLM rankings and leaderboards.

“We believe AI should serve as an extension of individual agency and, in the spirit of freedom, be distributed as widely and equitably as possible,” Murati wrote. “We hope this vision resonates with those who share our commitment to advancing the field.”

Some of the PR statements made certainly sound like those an open-source startup or at least an open-weight model maker might make. Bloomberg’s Mark Gurman considering Thinking Machines Lab as a potential Apple acquisition target are pretty far-fetched and amaze me.

AI explainability is truly important in terms of trust, security and the accountability of these systems:

“Soon, we’ll also share our best science to help the research community better understand frontier AI systems,” said Murati. - Tweet.

• In addition to Meta Superintelligence Lab poaching several core members of OpenAI’s team, there’s a continued exodus here too:

What does “Open Science” even look like in the future of America? 🤔

• Open-science

• Rapid iteration

• Co-design and collaborative principles

• Focus on real-world deployments

• Frontier capabilities with multi-modal nativity

Since Murati launched her venture, Thinking Machines Lab has attracted some of her former co-workers at OpenAI, including John Schulman, Barret Zoph (CTO), and Luke Metz. These are some of the most senior people that were left at OpenAI. These are some of the humble ones. Who take the science of research seriously.

The combined experience of these founding researchers at Thinking Machines Lab is off the charts as young a field as Generative AI and LLMS are. As we have seen with the talent wars, experience is at a premium.

Voice and Ambient Computing with LLMs?

Read more

In recent times it appears academic papers are being used as a form of public relations for various ideological positions along the AI hype cycle spectrum. This isn’t really anything new but it is striking.

Recursive Self-improving AI (RSI)

A couple of papers caught my attention around this trend of RSI, when in mid May Google DeepMind announced AlphaEvolve. This also comes across in Anthropic claiming how much Claude Code is used to make Claude Code itself better, i.e. claims like Claude Code wrote 80% of its own code.

Read the AlphaEvolve Paper

While Google says its deployed algorithms discovered by AlphaEvolve across Google’s computing ecosystem, including our data centers, hardware and software. There are also other research labs claiming other kinds of aspects of RSI or frameworks for it. While far fetched by the usual audacious claims by DeepMind, Sam Altman’s Gentle Singularity blog is even more extreme. As if the prophecy of Venture Capitalists could justify the enormous capital being spent on a pattern matching system like the transformer architecture where token prediction are coached or imbued with names like “reasoning models” and “superhuman intelligence”.

Read more

History is a funny thing, on some days the goalposts are moved.

Read Levels of AGI paper

Papers on AGI of Note | Related 📑

• Sparks of AGI (2023)

• …

Read more

Good Morning,

• I want to go over some of the AI News worth knowing about before I share some thoughts about AGI.

• In my various Notes I’m slipping in various AI news updates, so follow me there too if you are into that sort of thing.

• Don’t forget I also have a community chat, that’s fairly active. I ask my readers provocative questions about AI.

While OpenAI’s for-profit pivot hangs in the balance, one 23- year old YouTuber has become an AGI Podcaster/Youtuber, and by the way I think this is great, and hilarious! is a super interesting AGI interviewer now: Watch: Mark Zuckerberg – Meta's AGI Plan.

AGI is trending. 🚀 (And I’m not entirely sure how we should feel about this)

Leave a comment

Meta goes all-in on AI and Wearables

The reason this is interesting is because while Microsoft and Amazon appear to have scaled back AI Infrastructure plans, Meta has increased them. They just did earnings so let me explain.

• Meta just launched a stand-alone AI app to compete with ChatGPT

Meta AI App: Learn More

• Meta increased its 2025 capital expenditures to come in between the range of $64 billion to $72 billion due in part to “an increase in the expected cost of infrastructure hardware.”

• Microsoft’s Azure Cloud growth saw monster 33% growth. This for them is assurance that their AI Infrastructure spend (also called “capex”) is worth it.

This means that AI capex is not showing signs of slowing down in the race for AGI in the first half of 2025. This has big implications even if we are heading into a recession. Tariffs and trade wars appear to increase AI capex significantly or by about 11.2% approximately so far (going by Meta’s adjustment)!

What we do know is Meta will pivot to try to be an AI wearables company too, as demand for their smart glasses was higher than expected. Most of their so-called AI users are just on WhatsApp or Instagram maybe using some Generative AI features in passing. Now with a dedicated AI app and more AI wearables, things are about to get interesting for Meta’s so-called AI pivot (post Metaverse).

Alibaba Cloud’s Qwen 3

• Lost as usual in the Western AI news cycle are important Chinese open-source releases, Qwen 3 is an epic announcement.

Qwen3 Read More

• Their flagship model is a mouthful: Qwen3-30B-A3B

Basically it’s one of the best open-weight models out of China. They claim that Qwen3-235B-A22B, achieves competitive results in benchmark evaluations of coding, math, general capabilities, etc., when compared to other top-tier models such as DeepSeek-R1, o1, o3-mini, Grok-3, and Gemini-2.5-Pro.

While some uncertainty exists around Nvidia; Google, Meta and Microsoft all beat on Earnings as expected meaning for datacenters it’s likely all speed ahead even with more expenses added on due to semiconductor supply-chain disruptions due to trade wars.

Share

My Favorite Quote of the Week

One of my favorite quotes this past week is Nvidia’s CEO saying China is not behind in AI. I’ve been thinking about this due to Huawe’s new AI chip which I covered here. Nvidia CEO Jensen Huang said that “China is not behind” in artificial intelligence, and that Huawei is “one of the most formidable technology companies in the world.”

Speaking to reporters at a tech conference in Washington, D.C., Huang said China may be “right behind” the U.S. for now, but it’s a narrow gap. “We are very close,” he said. “Remember this is a long-term, infinite race.”

Huawei has started delivering its CloudMatrix 384 AI Clusters to customers already. Nvidia’s prime challenger might not be AMD, but actually Huawei in the 2030s, or even earlier.

OpenAI, Meta and Anthropic claims AI Supremacy like Future revenues

Here is OpenAI’s expected revenue trajectory. We can expect many new products in 2026 it seems.

Also while OpenAI is making odd claims about future revenue (see above), so is Meta: Meta forecasted it would make $1.4T in revenue from generative AI by 2035. AGI isn’t just a MAGA moonshot, it’s how BigTech keeps in power.

Soul-Searching our way to AGI

While BigTech and AI startups claim AGI soon, the way we are actually using1 Generative AI might surprise you.

People aren’t using ChatGPT for mind blowing enhancements of their intelligence, but rather more for personal problem solving and mental health interventions in 2025. Soul-Searching our way to AGI means data harvesting2 of our mental health and psychological profiles by companies like OpenAI and Anthropic. This will lead to advertising revenue for them later down the line.

How we use ChatGPT (or Claude or Gemini for that matter) differently is changing quarter to quarter a lot, so everything is still very much in flux. [see above graphic and HBR article here].

AI News in a Nutshell - in brief: 4:26

Where to go Next to Read about AGI?

Further Reading on AgI

Articles on AGI related themes recently around the Stack:

• 🌻 agi (disambiguation) by @jasmine Sun

• Agents are here, but a world with AGI is still hard to imagine by

  Harry Law

• AGI versus “broad, shallow intelligence” by Gary Marcus

• Our first project: AI 2027 by Daniel Kokotajlo

• Why I’m writing a Book about AGI by

  Kevin Roose

• Introducing AI 2027 by Scott Alexander (also a LessWrong graduate bro)

• And honestly, so many more.

😕 Event disclosure: This AGI piece was my humble submission to The Substack Post for their Unstacked Series. Unfortunately it didn’t make the cut.

• Though you might want to read this.

• The complete AGI breakdown for beginners

• How will we know when we have arrived at AGI?

• What AI writers on Substack are saying about AGI.

Subscribe now

Read more

Today’s guest post is an important deep dive into OpenAI’s Conversion Strategy of its business structure. At a later date it will be published as a stand-alone piece, without the context of this introduction.

OpenAI has raised a total of approximately $57.9 billion in funding over 11 rounds since its founding in December 2015. This makes it the most heavily funded AI company in the world with a significant centralization of power.

OpenAI has been called a generational company1, but what will the ChatGPT Titan become exactly? What will it need to do, even just to survive2.

The rivalry between Elon Musk and Sam Altman and the companies we know today as OpenAI and xAI might grow into being fairly dangerous clash for the future of artificial intelligence. There are some legal and business affairs I want you all to understand more deeply.

OpenAI is working on a Twitter Clone

Guys this is not a meme, or an exercise in napkin math, in late March, 2025 Elon Musk said X, formerly Twitter, had been acquired by xAI. On April 15th, the Verge broke the story that OpenAI is working on its own X-like Social network.

Aggressive scaling and acquisition plans

On April 16, Bloomberg fist reported that OpenAI is in talks to acquire Windsurf (formerly Codeium) for $3 Billion, a notable “vibe coding” platform. Also on April 16th, OpenAI released o3 and o4-mini. These new reasoning models can “agentically use and combine every tool within ChatGPT—this includes searching the web”, etc…

Clashes of the titans of America

Elon Musk and Sam Altman are now billionaires with some of the biggest funding capabilities of technology startups or AI startups, in history. What OpenAI is doing in AI Infrastructure alone is unprecedented. But OpenAI is virtually unrecognizable from the AI startup it used to be. What they will build could be powerful enough to disrupt the magnificent seven, BigTech in Silicon Valley as we know them, outside of Google’s stellar suite of AI products and leading large language model efforts. But the perversity OpenAI’s evolution is also mirrored in complicated legal issues and controversy of prioritizing commercial interests over the “benefitting of all of humanity.”

OpenAI’s race to change its Business Structure

The stakes are high for OpenAI, which needs to complete its for-profit conversion by the end of this year or next, or it will risk relinquishing some of the capital it has raised in recent months. It’s the race to alter its entire Mission. You might recall, OpenAI was founded as a non-profit in 2015. It turns out, a lot can change in a decade and ChatGPT was only released approximately 2.5 years ago.

This new structure will give the for-profit arm of OpenAI much more control. Under the proposed structure, the public benefit corporation, which is a for-profit corporate entity, will run and control OpenAI’s operations and business.

• On April 11th, a group of ex-OpenAI employees on filed a proposed amicus brief in support of Elon Musk in his lawsuit against OpenAI, opposing the company’s planned conversion from a nonprofit to a for-profit corporation.

“Generative AI as a massively disruptive sector required an unprecedented amount of proprietary resources and with it came a centralization of power that is the OpenAI we know today along with a “180 on all things AI Safety.”

• According to the Financial Times, OpenAI gave some testers less than a week for safety checks for an upcoming major launch.

• An organization OpenAI frequently partners with to probe the capabilities of its AI models and evaluate them for safety, Metr, suggests that it wasn’t given much time to test o3.

• Another of OpenAI’s third-party evaluation partners, Apollo Research, also observed deceptive behavior from o3 and the company’s other new model, o4-mini.

In fact, OpenAI updated its safety framework—and no longer sees mass manipulation and disinformation as a critical risks worth negating.

Read o3 and o4-mini System Card

• The exodus of AI governance, alignment and policy staff (talent) at OpenAI was a clear break from its past ideology with potentially irrevocable consequences. American corporations are conducting themselves with high-risk policies according to AI governance experts in an AI arms-race to dominate the future.

But as layoffs and entry level software engineers jobs dry up, is automating the future of coders a healthy goal in the first place?

Leave a comment

“Some of the most powerful and well-funded organizations in history are trying to build systems that would make humanity obsolete. It's far from clear that they will succeed, but regardless, these companies and the people running them should be critically examined as part of a broader project to collectively understand and govern the development of advanced artificial intelligence.” - Garrison Lovely (in about section of Obsolete)

The Betrayal of its original Mission

The brief, filed by Harvard law professor and Creative Commons founder Lawrence Lessig, names 12 former OpenAI employees: Steven Adler, Rosemary Campbell, Neil Chowdhury, Jacob Hilton, Daniel Kokotajlo, Gretchen Krueger, Todor Markov, Richard Ngo, Girish Sastry, William Saunders, Carrol Wainwright, and Jeffrey Wu.

“If the OpenAI Nonprofit agreed to a change in the OpenAI corporate structure which took away its controlling role, that would fundamentally violate its mission,” a filing on last Friday said.

In a nutshell:

OpenAI was founded as a nonprofit in 2015, but it converted to a “capped-profit” in 2019, and is now trying to restructure once more into a PBC. When it transitioned to a capped-profit, OpenAI retained its nonprofit wing, which currently has a controlling stake in the organization’s corporate arm.

OpenAI: "Our Structure"

• Musk’s suit against OpenAI accuses the startup of abandoning its nonprofit mission.

• Musk had sought a preliminary injunction to halt OpenAI’s conversion.

• I propose that what OpenAI and xAI becomes could make a lot of the internet we know today, obsolete.

In some ways we are already seeing evidence of this. (e.g. Elon Musk weaponizing X for political agendas, etc…)

To decode all of this let’s turn to investigative journalist and whistleblower .

Garrison is the author of the forthcoming book called

Obsolete: Power, Profit, and the Race to Build Machine Superintelligence.

He once wrote an article Can Humanity Survive AI, but I wonder if it can survive OpenAI3. Silicon Valley has a way of breaking you down and building you up into the venture capitalist mould and highly mutable (think opportunistic) Tycoon that Sam Altman has become.

Subscribe to Obsolete

Obsolete Newsletter does reporting and analysis on capitalism, the great power competition and the race to build machine superintelligence. File this one under our AGI section and OpenAI.

Garrison explains:

Obsolete 💾

Obsolete

Reporting and analysis on capitalism, great power competition, and the race to build machine superintelligence from journalist w/ work in NYT, Nature, BBC, TIME, and more.

By Garrison Lovely

Garrison has written about OpenAI before, here, here and here.

Read the Ex-OpenAI Amicus Brief

Given the extensive research he’s been doing for his book, I consider him an expert on OpenAI’s history and legal issues and the crucial months ahead for the company. When your Head of Preparedness steps down, yep we do take notice! Google sees itself hiring a “post-AGI” research scientist so who knows what is going on at this rate. The early idealism of Generative AI of the 2015 to 2020 period, has already nearly entirely faded as OpenAI and Anthropic are now the new duopology of AI in the United States that will bag most of the revenue from the hype cycle.

"The Profit Motive is Winning": OpenAI's Philanthropic Gambit Faces Criticism

Here is a key quote from the brief:

“OpenAI committed to several key principles for executing on [its] mission in their charter document,” the brief reads. “These commitments were taken extremely seriously within the company and were repeatedly communicated and treated internally as being binding. The court should recognize that maintaining the nonprofit’s governance is essential to preserving OpenAI’s unique structure, which was designed to ensure that artificial general intelligence benefits humanity rather than serving narrow financial interests.”

Garrison Lovely4 is not an ordinary freelance journalist, he’s a Reporter in Residence at the Omidyar Network. He lives in Brooklyn and in addition to covering AI, he’s written about psychedelics, prisons, emerging technologies, philosophy, fish welfare, management consulting, police alternatives, and more. He’s also conducted interviews with Michael Stipe of REM and Conor Oberst of Bright Eyes. He has an exceedingly strong grounding in investigative journalism, so his reporting is like looking into a crystal lake.

Listen to his recent Talk

With that, let’s get into Garrison’s deep dive:

(I really lucked out getting Garrison to agree to share this piece with us):

TL:DR try the audio version: 23:04

Inside OpenAI's Controversial Plan to Abandon Its Nonprofit Roots

By , April 16th, 2025.

About the author

Former employees, legal experts, and philanthropic leaders challenge the company's effort to shed nonprofit control

Earlier this month, OpenAI announced that it aspires to build "the best-equipped nonprofit the world has ever seen" and was convening a commission to help determine how to use its "potentially historic financial resources."

But critics view this new commission as a transparent attempt to placate opposition to its controversial plan to restructure fully as a for-profit — one that fails to address the fundamental legal issues at stake.

OpenAI is currently a $300 billion for-profit company governed by a nonprofit board. However, after an earlier iteration of that board briefly fired CEO Sam Altman in November 2023, investors reportedly began demanding that the company shed its quasi-nonprofit status.

"The story of OpenAI's history is trying to balance the desires to raise capital and build the tech and stay true to its mission," a former OpenAI employee told me. The current move, they say, is an attempt to "separate these things" into a purely commercial entity focused on profit and tech, alongside a separate entity doing "altruistic philanthropic stuff."

"That's wild on a number of levels because the entire philanthropic theory of change here was: we're going to put guardrails on profit motives so we can develop this tech safely," the former employee says.

Best newsletter pieces - Obsolete

• “End of an Era”: OpenAI’s AGI Readiness Chief Resigns and Team is Disbanded

• Is Deep Learning Actually Hitting a Wall?

• Why Did Elon Musk Just Offer to Buy Control of OpenAI for $100 Billion?

• What the Headlines Miss About the Latest Decision in the Musk vs. OpenAI Lawsuit

• Is the AI Doomsday Narrative the Product of a Big Tech Conspiracy?

Legal hurdles

The for-profit conversion faces significant unresolved legal challenges, including a lawsuit from Elon Musk arguing that his $44 million donation was contingent on OpenAI remaining a nonprofit and that the conversion would violate its founding charitable purpose. The case will go to trial this fall. The conversion can also be challenged by the California and Delaware Attorneys General (AGs), who are reportedly each looking into the case.

Historical Timeline:

Created in Datawrapper, April 17th, 2025.

Musk's suit, OpenAI's gargantuan valuation, and the unprecedented nature of the conversion attempt have attracted scrutiny.

Without mentioning OpenAI explicitly, California Assembly Member Diane Papan introduced a bill in February that would have blocked the conversion. However, the legislation was amended without explanation earlier this month to instead focus on aircraft liens. Papan's office has not replied to a request for comment.

Read more

Subscribe now

Welcome Back,

And welcome to our next article in our series on AGI.

We start off with a simple question, will agents lead us to AGI? OpenAI conceptualized agents as stage 3 of 5. You can ascertain that agents in 2025 are barely functional.

Since ChatGPT was launched nearly 2.5 years ago, outside of DeepSeek, we haven’t really seen a killer-app emerge. It’s hard to know what to make of Manus AI? Part Claude wrapper, but also an incredible UX with Qwen reasoning integration. Manus AI, which has offices in Beijing and Wuhan and is part of Beijing Butterfly Effect Technology. The startup is Tencent backed, and with deep Qwen integration you have to imagine Alibaba might end up acquiring it.

Today technology and AI historian, Harry Law of , explores this awkward stage we are at halfway between reasoning models and agents. This idea that agents will lead to AGI is also quite baffling. You might also want to read some articles of the community on Manus AI: but will “unfathomable geniuses” really escape today’s frontier models, suddenly appearing like sentiment boogeymen saluting us in their made-up languages?

Explore our AGI Series

Read about Manus AI

1. Manus: China’s Latest AI Sensation (China Talk)

2. The Manus Marketing Madness (Don’t Worry about the Vase)

3. Why Manus Matters (Hyperdimensional)

One thing quickly becomes clear, Manus AI is about discovering and building the best interface and UX for agents, not about AGI. Manus AI is about more than being just a chatbot that spits out good answers or pseudo-reasoning models capable of some independent exploration. Manus has called itself: the world’s first autonomous agent.

New Qwen Release on March 26th, 2025

Qwen (Tongyi Qianwen) AI is developed by Alibaba Cloud, the cloud computing division of Alibaba Group, which is a major player in e-commerce and technology in China.

Qwen2.5-Omni was just released, that includes a “Thinker-Talker architecture” enabling Voice Chat + Video Chat! Just in Qwen Chat (https://chat.qwen.ai). I have no affiliation with Qwen but I do own Alibaba ADR stock (in very small amounts). Overall I think Qwen is under-appreciated in the West as a technical peer of DeepSeek, and maybe even a bit ahead. It’s benchmark performance are worth understanding.

Bio

Harry Law writes Learning From Examples, a history newsletter about the future. AI is a major focus, but he writes about any episode that helps us think more clearly about technological change. He’s a PhD candidate at the University of Cambridge and a former researcher at Google DeepMind.

His academic pursuits primarily focus on the history, philosophy, and governance of AI, which situates him at the intersection of technology and ethics. I really enjoy his essays, and I hope you do too!

Examples of recent articles

• The Slop Must Flow

• A Verhoevenian Moment

• Bovril Among the Machines

Tl;dr Audio Version (I’ve slowed down the narration to help us fully process this incredible essay)

By Harry Law, March, 2025.

Agents are here, but a world with AGI is still hard to imagine

Are you feeling the AGI? For a long time, it was taboo in the ‘serious’ AI research world to even mention it. Then AlexNet happened, then AlphaGo, then the transformer. By the time ChatGPT gave the public a taste of the ‘unreasonable effectiveness’ of large models, labs were openly discussing AGI as part of their product roadmap like it was just another feature to be rolled out, tested, and debugged.

Now even journalists, who generally think AGI is a sinister marketing ploy, are starting to think otherwise. Kevin Roose reckons it’s just around the corner and so does Ezra Klein. But if AGI is getting closer, shouldn’t we have a better idea of what it will actually look like?

For the past few years, there’s been a lot of talk about the transition from the current ‘tool-based’ paradigm to what developers call the ‘agent-based’ paradigm. If tool-based AI is about using prompts that require repeated human input, then agent-based systems can take action autonomously and act without it.

This sounds clear enough, but we should remember that autonomy is a spectrum. Since just after the launch of ChatGPT, indie developers released ‘agentic’ models like BabyAGI and AutoGPT –– but they generally weren’t reliable enough to go anywhere. Hobbyist efforts like these, which work through automatic prompting, were followed by projects from firms with mountains of venture capital cash like coding agent Devin.

Reviews for these sorts of systems are generally mixed. Some report lots of success with Devin and agents like it, others found they didn’t quite meet their expectations (in many ways the story of the large model era). Others wondered whether Devin, which was built atop of OpenAI’s GPT models, would soon be superseded by an effort from a major lab. New work suggests that the length of tasks that AI systems can do is doubling about every 7 months, which is a solid yardstick for measuring agentic progress.

Enter Manus

China’s Manus AI isn’t the first AI agent of the large model era. We have the likes of Operator and Deep Research (OpenAI), Computer Use (Anthropic), and Project Astra (Google DeepMind). What it is, though, is the first generalist agent to enter the public imagination in a meaningful way.

There’s a strangeness to the Manus episode. We’re not looking at a ChatGPT-level moment––and probably not even a DeepSeek-level one––but it may give us a very small hint about what a world with AGI could look like. That we can make that claim at all is surprising given how long AI watchers have been expecting agents to appear en masse.

Manus is billed as ‘a general AI agent that bridges minds and actions’. In what I take as a jab at the tool-based paradigm, its developer says that Manus ‘doesn't just think, it delivers results’. Similar to efforts from OpenAI and Anthropic, Manus is basically like giving someone else access to your computer remotely. It can research and write a nice one-page memo, build a promotional network of social media accounts or book tickets and hotels for a conference.

Like Devin, the agent is built on top of existing models –– in this instance the much loved Claude 3.5 Sonnet from Anthropic and a fine-tuned version of Alibaba's open-source model Qwen. This ensemble approach is both significant (in that it seems to generate some promising results) and unsurprising (in that this kind of approach has been imagined by AI researchers since at least the 1980s).

Marvin Minsky, one of the grandees of the AI project, said that he thought the brain was what computer scientists call a ‘kludge’. He believed that it was an inelegant solution to the challenges faced by early humans, cobbled together from specialised parts over the course of millennia. The amygdala is central to processing emotions. The hypothalamus helps regulate basic bodily functions. And the hippocampus plays a crucial role in memory formation.

Like the brain, I suspect anything that’s generally thought to be AGI will be the product of both specific mechanisms and their patterns of interaction. As Minksy imagined, we can probably look forward to lots of small, dedicated models that can be linked together as part of a more complex whole.

That isn’t to say that Manus is anywhere close to AGI, but rather that it may share some conceptual similarities with something that eventually does. This is why I think questions about whether LLMs will scale all the way to AGI completely miss the point: they don’t need to.

The Elephant in the Chinese Room

Manus was developed by a Chinese firm. It comes hot on the heels of DeepSeek’s R1 reasoning model, whose release was described as a Sputnik moment for American AI by Marc Andreessen.

Quibbles about the accuracy of that particular parallel aside, it is true that American AI labs have been betting the house that bigger models mean better models since the release of ChatGPT. This is still the case, but DeepSeek-R1 confirms that there is another option open to developers: instead of using more compute to make the underlying model bigger, they can spend it on improving the quality of outputs as they happen.

DeepSeek-R1 works by creating a chain of responses to mimic the process of thinking out loud. At each stage it picks the best option before proceeding on to the next, which results in a much stronger final output. It’s the difference between writing the first thing that comes into your head during an exam, and planning your essay before you begin.

The significance of DeepSeek’s work was two-fold. First, the group found a way to make the process of building the underlying model much cheaper. Although the $5M price tag dramatically understates the true cost by ignoring spending on labour and hardware, the model still represents a major improvement in efficiency. Second, DeepSeek was able to get to market quickly and share an open source version of its model.

Much like Manus, the underlying approach that made DeepSeek-R1 a success wasn’t new. Chain of thought reasoning, like computer use functionality, had been up-and-running inside American AI labs for quite some time. So how is it that the Chinese lab seems to have released the first publicly available model first?

It comes down to risk appetite. If you use Manus, you’ll discover that it isn't perfect. It’s a remarkable piece of software capable of some very useful things, but it does make mistakes. When a model is capable of taking over your computer and losing you or someone else money, then you’re naturally going to think twice about releasing it to the public.

The US AI labs are, for obvious reasons, keen to avoid releasing a model that could potentially cause significant damage. They have much more to lose than the folks behind Manus, who could only dream of a trillion dollar market cap.

But dream they do. Team Manus is happy to take a calculated bet: that its flagship model can be released to a sufficiently small number of users over time, and that feedback from these users can be tapped to make it less likely to cause mayhem in the future.

Agents, broad and narrow

The most well known agent today is probably Devin. It’s an example of a ‘narrow’ agent that works specifically on coding. No one would compare it to AGI because it can’t perform anywhere near the number of tasks a human can.

OpenAI’s most general agent is Operator, which like Manus exists as a browser-based product that can perform many types of tasks (see a direct comparison here). Manus is generally credited with being capable of doing more at a higher quality, and is pitched as a slightly more polished article versus the Operator ‘research preview’. Anthropic likewise has its Computer Use agent, which is also in beta testing and can only be accessed via API.

The most slick agent from the major American labs is OpenAI’s Deep Research, which can perform research tasks like conducting literature reviews and analysing sources. I find Deep Research to be useful in my work when I’m looking for a rough start on research tasks, but I understand others' frustration with it. Like many of the frontier AI products, whether you find it helpful comes down to a combination of the way you use it, what you use it for, and the expectations that you have.

Deep Research costs a hefty $200 a month, but that’s pocket change compared to the figure reported by some outlets for OpenAI’s next products. According to The Information, OpenAI intends to launch a suite of agent products for specific applications like sales or software engineering.

A ‘high-income knowledge worker’ agent could cost $2,000 a month, while a software developer agent is reported to be priced at $10,000 a month. A PhD-level research agent may be up to $20,000, which is probably more expensive than employing a good chunk of humans in the same role.

For now, though, most narrow agents are as reliable as they are expensive (that is to say, ‘kind of’). It’s also self-evident that today’s specialised agents don’t feel like AGI. I’m not necessarily sure that Manus does either, but it does at least feel closer to what I imagine AGI will actually be like. The clue is in the name. Manus is a general agent, a jack of all trades. Yes, it’s worse than Deep Research at writing reports –– but it can take over my computer and book a hotel.

Despite failing every now and then, it’s generally useful in a way that other agents haven’t been. Factoring in grip on the discourse, I think Manus is a good foil for thinking about the future. I expect we will probably have much more robust systems that can conduct a wider range of actions sooner than we think.

The labs like to talk about agents as drop-in remote workers who can do anything that an employee can. I think despite the obvious limitations––the basic errors, the lack of complete multimodal functionality, and issues with staying on task over time––it certainly seems like a window into the near future.

But that preview of tomorrow is only useful as it relates to a person deploying a single agent. A world with AGI isn’t just about you managing one agent to do a limited number of tasks; it’s one where you could manage hundreds of them simultaneously. Or––as is more likely––you might oversee one or a handful, which in turn look after another tranche of instances. And all the way down it goes within your own personal AI organisation. Manus doesn’t really tell us much about how that dynamic will manifest.

Another area of the post-AGI universe shrouded in the fog of war is the second and third order effects on society at large. If everyone has huge numbers of agents, you will inevitably be interacting with them on a day-to-day basis (or at least your agents will be dealing with the agents of others). This is to say nothing about the macroeconomic implications of AGI, which may change the world very quickly even assuming a relatively modest impact on GDP growth.

It may be that some of these effects are drawn into focus as systems get increasingly powerful. Most AI people I know are adamant that AGI development is likely to be continuous rather than discontinuous. That is to say, there will be no single moment in time when AGI happens. No one is going to roll it out from behind the curtain to a gasping audience.

Instead, we’ll build better models, deploy them, and modify the next generation based on how they perform in the real world. This is basically what all of the major labs do in one form or another –– and in some respects is the approach taken by Manus (in the most maximalist version of this strategy).

Calling AGI will be like boiling a frog. We’ll become blasé about the capabilities of the systems of the future, about how they are changing the world around us in ways that were previously thought impossible. Some people already think AGI exists. In a few months, perhaps more people will too. Even more in a couple of years.

Eventually, the world will be turned on its head and those saying AGI isn’t here will be in the minority. Manus probably hasn’t changed much in that respect, but it does represent another small step in that direction.

By Harry Law

Learning From Examples

A history newsletter about the future

By Harry Law

Harry’s essays are super historical, refreshing and philosophical reading. I’m not usually a fan of reading AI policy and governance types, but Harry is cut from another cloth altogether. [Doing a PhD at the Department of History and Philosophy of Science]. Harry infuses a historical context into his writing which makes it truly come alive!

More by Harry

• Reflections on AGI from 1879

• Walden Three: AI and the intention economy

• The Economy of Magic

Welcome Back,

Everyone from OpenAI to DeepSeek claims they are an AGI startup, but the way these AI startups are proliferating is starting to get out of control in 2025. I asked Futuristic Lawyer , to look into this trend.

On 14 April 2023, High-Flyer announced the start of an artificial general intelligence lab dedicated to research developing AI tools separate from High-Flyer's financial business. Incorporated on 17 July 2023, with High-Flyer as the investor and backer, the lab became its own company, DeepSeek.

But while saying you are an AGI research lab has come into popular fashion in marketing terms in recent years, does anyone even believe AGI is a real thing or that today’s architecture even has the capability of attaining it?

The definition of and the date when it is achieved are both hotly debated. However it seems actual machine learning engineers and researchers don’t actually think the current LLM architecture can reach this apparent goal.

How AI Researchers view AGI

According to a recent survey of 475 AI researchers by the Association for the Advancement of Artificial Intelligence (AAAI) conducted as part of its panel on the future of AI research found that “[t]he majority of respondents (76%) assert that ‘scaling up current AI approaches’ to yield AGI is ‘unlikely’ or ‘very unlikely’ to succeed, suggesting doubts about whether current machine learning paradigms are sufficient for achieving general intelligence.”

Read the Report

The “Creating Einstein in a Datacenter” Problem

Maxwell Zeff of TechCrunch recently wrote a great analysis of this. In a piece this month, Thomas Wolf, Hugging Face’s co-founder and chief science officer, called some parts of Amodei’s vision “wishful thinking at best.”

Read the Short Essay on X

In Wolf’s opinion shared by Yann LeCun (and too many others to name), today’s LLMs simply aren’t up to the task (of AGI).

“I am not interested anymore in LLMs. They are just token generators and those are limited because tokens are in discrete space. I am more interested in next-gen model architectures, that should be able to do 4 things: understand physical world, have persistent memory and ultimately be more capable to plan and reason.” - Yann LeCun, Nvidia GTC, 2025

However comparing models to PhDs (which is absurd on the face of it) to promises of AGI aren’t stopping AI startups from raising huge funding rounds. Silicon Valley is in the business of sales, and Sam Altman is like a drunken sailor of false promises. This grift to Venture Funds and the public hasn’t just been a kind of Silicon Valley hallucination or hoax, it’s led to BigTech spending tens of Billions on AI Infrastructure.

It’s shaping up to be a very expensive narrative.

“To create an Einstein in a data center, we don’t just need a system that knows all the answers, but rather one that can ask questions nobody else has thought of or dared to ask.” - Thomas Wolf, Hugging Face.

The AGI Startups

The spin-offs from OpenAI seem to be more keen to wear the AGI mantle.

The recent statements of Anthropic’s CEO have been especially concerning. While OpenAI itself might be the ultimate pretender, pretending that you are building something essential that enables AGI to take place is getting fairly popular in 2024 and 2025 as well:

• We now have a legion of so-called AGI startups

Generative AI investment reached over $56 billion in venture capital funding alone in 2024. How much will it reach in 2025? How much will Capex for datacenters and AI Infrastructure increase in the frantic years ahead?

My Proposed List of AGI Startups

The person at Microsoft who wrote the “Sparks of AGI” paper even ended up joining OpenAI. Microsoft first investing vast sums of money in OpenAI in April, 2023 signaled the start of a new kind of grifting culture in Silicon Valley, that has even spread to China in 2025.

1. Anthropic by seven co-founders (now Billionaires)

2. xAI by Elon Musk ⭐

3. Safe SuperIntelligence (SSI) by Ilya Sutskever ⭐

4. Thinking Machines Lab by Mira Murati ⭐

5. Ndea by François Chollet ⭐

6. DeepSeek by Liang Wenfeng

7. Reflection AI by Misha Laskin

8. Moonshot AI by Yang Zhilin and Zhang Yutao

9. Zhipu AI by Tang Jie and Li Juanzi

Care to estimate the total amount in U.S. dollars the above startups are going to raise in their lifetime? It’s going to be a staggering amount. These are the research labs that I most consider aligned with being an “AGI startup” as of March, 2025. They are likely to keep multiplying, some of them even taking robotics form, e.g. Generalist AI.

These companies aren’t building machines of loving grace and their efforts are very likely going to waste a lot of valuable capital, human talent, and time. Dario Amodei, the CEO and cofounder of Anthropic recently claimed AI will write 90% of code in 3-6 months and nearly all code within a year, potentially transforming software development that impacts the industry significantly.

This is clearly serious business! The founder of Reflection AI didn’t think AGI was even good enough, they called their startup a Superintelligence startup, previously only dared by Ilya Sutskever himself. Thinking Machines Lab by Mira Murati in fact is so serious, it needs dozens of former OpenAI employees to get off the ground running. After the commercial success of Anthropic (the first spin-off of the OpenAI Mafia), anything goes! 🚀

Futuristic Lawyer

Futuristic Lawyer

Making sense of grey areas between law, IT and ethics.

By Tobias Mark Jensen

On Futuristic Lawyer Newsletter, Tobias Jensen writes about the knowledge gap between big tech companies and democratic institutions.

"The vast investments in scaling, unaccompanied by any comparable efforts to understand what was going on, always seemed to me to be misplaced," Stuart Russel, a computer scientist at UC Berkeley who helped organize the report, told New Scientist.

BigTech and now European and Chinese capex in AI Infrastructure and the funding of these “AGI startups” is starting to get out of control in 2025. Even when there is consensus that scaling up LLMs won’t lead to anything approaching AGI.

François Chollet himself a founder on our list with Ndea, as himself argued that while AI might be capable of memorizing reasoning patterns, it’s unlikely it can generate “new reasoning” based on novel situations. Just don’t tell that to Japanese based Sakana AI. These young prodigies from either OpenAI or Google, usually have commercial interests in pretending otherwise. Ndea itself plans to use a technique called program synthesis, in tandem with other technical approaches, to unlock AGI.

Articles by Tobias Jensen

• AI Could be Heading Towards the Trough of Disillusionment

• How to Deal with Data Harvesting AI Girlfriends?

• China's Position in AI & BigTech, an Interview with

In this article, Tobias Jensen does a deep dive on some of the AGI startups worth mentioning in 2025:

For less than $2 a week unlock our full coverage.

The New Generation of AGI Startups

By Tobias Jensen, March, 2025.

We are seeing a new generation of AGI startups emerge into the market. They are not racing to compete with the capital expenditures of OpenAI, Microsoft, Meta, Google, Anthropic, x.AI, and others. Instead, they are reimagining whole new approaches to AI development and exploring alternative pathways to AGI.

See other pieces I have written about AGI on AI Supremacy here:

• Scaling Barriers, Diminishing Returns and the Great AI Ceiling of Over-Promises

• The Rise of the AGI Profiteers

• GPT-4's Sparking Intelligence

Three AGI Startups

In today’s guest post, we will take a look at three notable AGI startups: Safe Superintelligence (SSI), Thinking Machines Lab, and Ndea.

Each of these is co-founded by former top researchers at the major labs in the US, Ilya Sutskever, OpenAI’s former Chief Scientist, Mira Murati, OpenAI’s former Chief Technology Officer, and François Chollet, former software engineer and deep learning researcher at Google, respectively.

At least on paper, Sutskever, Murati, and Chollet held very attractive positions at the world’s leading AI labs. So - why did they decide to leave their well-paying day jobs and pursue ventures of their own?

For different reasons of course, but there seems to be a common thread.

Ilya Sutskever told Reuters in November 2024:

"The 2010s were the age of scaling. Now we're back in the age of wonder and discovery once again. Everyone is looking for the next thing. Scaling the right thing matters more now than ever."

Mira Murati has not shared much about her reasons for leaving OpenAI but she did say:

“I want to create the time and space to do my own exploration.”

During an epic episode of Dwarkesh Patel’s podcast, François Chollet made several statements about the current state of AI which also reveals his motivation for founding Ndea:

“if you scale up the size of your database and you cram into it more knowledge, more patterns and so on, you are going to be increasing its performance as measured by a memorization benchmark. That's kind of obvious. But as you're doing it, you are not increasing the intelligence of the system one bit. You are increasing the skill of the system. You are increasing its usefulness, its scope of applicability, but not its intelligence because skill is not intelligence. And that's the fundamental confusion that people run into is that they're confusing skill and intelligence.”

Later in the episode, Chollet says:

“I think OpenAI basically set back progress towards AGI by quite a few years, probably like five to ten years, for two reasons. And one is that, well, they caused this complete closing down of research, frontier research publishing. But also, they triggered this initial burst of hype around LLMs.

And now LLMs have sucked the oxygen out of the room. Like everything, everyone is just doing LLMs. And I see LLMs as more of an off-ramp on the path to AGR, actually. And all these new resources, they're actually going to LLMs instead of everything else they could be going to.

if you look further into the past to like 2015, 2016, there were like a thousand times fewer people doing AI back then. And yet I feel like the rate of progress was higher because people are exploring more directions. the world felt more open-ended.”

Sutskever, Murati, and Chollet are all expressing a desire to explore new approaches to AI development. Although Chollet is the only one who dares to speak out on record, my guess is they all have identified the same limitation. The pioneering AI labs display a tunnel-vision focus on increasing data, compute, and parameters to improve the capabilities of AI models. We call it the “scaling paradigm” or the “Bigger-is-Better Paradigm”. Sam Altman describes the idea in a recent blog post on his website about the economics of AI:

“The intelligence of an AI model roughly equals the log of the resources used to train and run it”.

The approach is conveniently simple for a non-technical person like Sam Altman to explain to investors. More money, more intelligence. But is OpenAI climbing the wrong ladder by focusing so much on scale, while neglecting entirely new directions of research?

A paper from September 2024 by Research Director at Inria Saclay Centre in Paris, Gael Varoquaux, AI and Climate Lead at Hugging Face, Alexandra Sasha Luccioni, and Signal CEO, Meredith Whittaker convincingly refute the Bigger-is-Better Paradigm with three main arguments:

1. The Bigger-is-Better Paradigm is not sustainable since “compute demands increase faster than model performance”. Concretely, performance on benchmarks tends to saturate on many tasks after a certain point.

2. The focus on scaling general-purpose AI stirs attention away from studying how the models really work which in turn hampers efforts to audit and evaluate models. Additionally, for many specialized applications (e.g. in health, education, or the climate) “utility does not require scale” meaning that smaller models can suffice.

3. The Bigger-is-Better Paradigm exacerbates a concentration of power in the hands of a few actors while threatening to disempower others in the context of shaping both AI research and its application throughout society.

DeepSeek ⭐

Most readers will recall how a certain Chinese startup took the internet by storm a few weeks ago by publishing astounding test results with their open-source reasoning model DeepSeek-R1, delivered on a very modest training budget. Should DeepSeek be included in the name pool of new AGI startups? Let’s briefly address that, before we move on to talking about SSI, Thinking Machines Lab, and Ndea.

The company’s CEO Liang Wenfeng has explicitly mentioned in interviews that DeepSeek is focused on “achieving AGI”, not on monetization strategies. However, this could change, and change fast. South China Morning Post (SCMP) reported on February 18 that the company DeepSeek has updated the scope of its business registry information to include “internet information services” - a move which could indicate plans to monetize its technology.

And why not? DeepSeek’s commercial potential is enormous. The company’s open-source AI models now provide AI assistance in devices by Chinese smartphone makers such as Huawei Technologies, Oppo, Vivo, Lenovo Group and Honor. Tencent Holdings and Baidu have integrated DeepSeek’s models into their search products. Chinese car markers including BYD, Geely, Great Wall Motor, Chery Automobile, and SAIC Motor have announced plans to use DeepSeek’s models in their vehicles. Even local government agencies are incorporating DeepSeek into their digital systems.

Much has been written about how DeepSeek outperformed America’s tech darlings on a shoestring budget. Nonetheless, semiconductor research and consulting firm SemiAnalysis estimates that DeepSeek has spent well over $500 million on computing hardware alone over the company’s history, excluding steep R&D costs and other preliminary costs that are not accounted for in DeepSeek-R1’s research paper.

All this evidence considered, I think of DeepSeek as a competitor to the major AI labs in the US, more so than a part of the new generation of AGI startups.

Safe Superintelligence (SSI) ⭐

Ilya Sutskever is among the most prominent names in the AI industry. He was a co-founder of OpenAI and worked as the company’s Research Director/Chief Scientist for more than 8 years. Sutskever was the main contributor to the architecture of OpenAI’s early GPT models, so I don’t think it’s an overexaggeration to say that he laid the foundation for OpenAI’s success.

On a personal level, Sutskever is well-known for possessing an esoteric belief in the potential of AI and the coming of a new “superintelligence”. According to a popular article by The Atlantic, Sutskever unironically led employees in a chant at an OpenAI company party: “Feel the AGI! Feel the AGI”.

During the chaos at OpenAI chaos in November 2023 when Altman was fired and rehired as CEO in the same week, Sutskever voted to remove Sam Altman as CEO, but then apologized on X afterward and signed an employee letter calling for the entire board to resign and for Altman to return. The course of events, which we will probably never know the full details of, was in all likelihood a decisive factor in Sutskever’s departure from the company. Additionally, Sutskever was heading OpenAI’s efforts on general AI safety research in the “superalignment team” which was officially disbanded soon after he left in May 2024.

Safe Super Intelligence (SSI) Inc. was founded in June 2024 by Sutskever with Daniel Levy, former OpenAI researcher, and Daniel Gross, former AI lead at Apple. In breaking with OpenAI’s commercially focused approach, SSI is not concerned about selling everyday AI products to consumers but on rapidly advancing “superintelligence” while emphasizing safety:

“We approach safety and capabilities in tandem, as technical problems to be solved through revolutionary engineering and scientific breakthroughs. We plan to advance capabilities as fast as possible while making sure our safety always remains ahead.

This way, we can scale in peace.

Our singular focus means no distraction by management overhead or product cycles, and our business model means safety, security, and progress are all insulated from short-term commercial pressures.”

Even before SSI has launched a product, a road map, or a business plan, the company has raised $1 billion in capital from investors, including Sequoia, DST Global and SV Angel. As reported by Bloomberg, SSI is at the time of writing close to raising another funding round of more than $1 billion at a $30 billion valuation.

Quite impressive for a company that isn’t generating any revenue nor have plans to do so in the near future.

Thinking Machines Lab ⭐

Mira Murati played a role in OpenAI’s growth journey from a small non-profit organization on a bizarre mission to its national positioning as a key player in America’s energy infrastructure, security, and economic development.

Murati joined OpenAI in 2018 as “VP of Applied AI & Partnerships” with prior experience from Tesla as Senior Product Manager. She took the role of Chief Technology Officer (CTO) at OpenAI in 2020 and had the chance to serve as interim CEO for a few days during Sam Altman’s brief ouster. Murati announced that she would leave OpenAI in September 2024, shortly after OpenAI released its new series of “reasoning models” with the launch of o1-preview and o1-mini.

On February 18, 2025, Thinking Machines Labs came out of stealth. Murati is CEO and co-founder but the team consists of more than a dozen former OpenAI employees, including the chief scientist John Schulman who co-founded OpenAI and worked for a brief stint as AI safety researcher at Anthropic, the CTO Barret Zoph who served as OpenAI VP of research, and Lillian Weng who used to work as OpenAI’s President of Safety.

Thinking Machines Labs is structured as a Public Benefit Corporation (PBC) - a company structure used by for-profit companies with a public-good mission. As a PBC, Thinking Machines Lab does not only have a fiduciary responsibility to maximize returns for shareholders but also to publish an annual “Benefit report” where the company outlines its progress towards its public benefit goals.

Thinking Machines Lab is on a mission to improve the scientific community’s understanding of frontier AI systems to make them “more widely understood, customizable and generally capable”. The company also emphasizes that "scientific progress is a collective effort” and that it wants to collaborate with the wider community. A clear deviation from OpenAI’s famous closedness which has become an industry norm.

“Knowledge of how these systems are trained is concentrated within the top research labs, limiting both the public discourse on AI and people's abilities to use AI effectively. And, despite their potential, these systems remain difficult for people to customize to their specific needs and values.”

Thinking Machine Labs is working to build “more flexible, adaptable, and personalized AI systems" that can “work with people collaboratively”. The company avoids using terms in its mission statement such as "superintelligence" and "AGI" but this is clearly what the founders believe in and are working towards. The mission is to create “advanced multimodal capabilities” and “ultimately (..) unlock the most transformative applications and benefits, such as enabling novel scientific discoveries and engineering breakthroughs.”

Overall, it sounds like Thinking Machine Labs is on the same mission as OpenAI but with an approach rooted in scientific collaboration and open research, rather than scaling.

“We'll focus on understanding how our systems create genuine value in the real world. The most important breakthroughs often come from rethinking our objectives, not just optimizing existing metrics.”

Ndea ⭐

François Chollet launched Ndea in January 2025 with his co-founder Mike Knoop, both have a background in software engineering. Chollet brings nearly ten years of work experience from Google, while Knoop is the co-founder and former Product and engineering lead at the software automation platform and industry giant, Zapier.

Chollet is well-known in the AI community for two major contributions. First, prior to joining Google in 2015, Chollet developed the popular deep learning development framework Keras which provides an interface for AI neural networks in the programming language Python. Second, Chollet published the paper On the Measure of Intelligence in November 2019 which critiques classical benchmark tests for failing to measure AI’s general intelligence. The paper was accompanied by the ARC-AGI benchmark (ARC-AGI: Abstraction and Reasoning Corpus for Artificial General Intelligence) which is intended as a more truthful proxy measure for AI intelligence (or AGI). The test consists of many puzzles that are deliberately designed to be easy for sharp humans and difficult for AIs to solve.

Chollet and Knoop launched the ARC Prize Foundation in June 2024, which offered prizes of $ 1 million in 2014 for the teams who could deliver the best results on the ARC-AGI benchmark with an open-source solution. The winning team scored 53.5% on the benchmark and won $25.000, while the best paper was awarded with a prize of $50.000. A new ARC prize competition is expected to launch as soon as Q1 2025 based on a new version of the benchmark called ARC-AGI-2.

According to Ndea’s website, the name - like 'idea' with an 'n' - is inspired by the Greek concepts ennoia (intuitive understanding) and dianoia (logical reasoning). In the same spirit as the ARC Prize Foundation, Ndea is dedicated to exploring how AI can develop genuine intelligence rather than skillful maneuvering of benchmark tests. Specifically, Ndea wants to combine deep learning with another promising but much less mature research field known as “program synthesis”:

“Instead of interpolating between data points in a continuous embedding space, program synthesis searches for discrete programs, or models, that perfectly explain observed data. This allows it to achieve much greater generalization power with extreme data-efficiency, requiring only a few examples to learn.”

The goal is to supercharge scientific advancements with AI. To achieve this purpose, Chollet and Knoop’s claim and bet is that current deep-learning AI cannot suffice because it “crumbles when faced with open-ended problems” and is “constrained by what humans teach it”. A more general form of intelligence is needed to truly take AI development to the next level, and the scaling paradigm is not enough.

“Today, the acceleration of scientific progress hinges on one factor: AI capable of independent invention and discovery. This capacity is the gateway to advancements beyond our wildest imagination.”

Wrapping Up

The new generation of AGI startups are branch-offs from the major AI labs. They are not fixated on quickly bringing new consumer products to the market or generating fast returns to justify enormous compute spendings. Instead, the AGI startups are focused on exploring new research directions so “AGI can benefit all of humanity” in line with OpenAI’s mission statement which is arguably abandoned by the company in practice.

Ilya Sutskever’s SSI is focused on AI safety¸ and secondly, on building advanced capabilities. Mira Murati’s Thinking Machines Lab is focused on collaborative research and increasing the scientific understanding of advanced models. François Chollet’s Ndea is focused on AI autonomy and the research direction “program synthesis”. If the AGI startups can attract the right people and succeed in new research directions, they could potentially raise existential and uncomfortable questions to the established AI industry, like DeepSeek did.

We wish them the best of luck.

Editor’s Notes

• Wolf thinks that AI labs are building what are essentially “very obedient students” — not scientific revolutionaries in any sense of the phrase.

“We're currently building very obedient students, not revolutionaries. This is perfect for today’s main goal in the field of creating great assistants and overly compliant helpers. But until we find a way to incentivize them to question their knowledge and propose ideas that potentially go against past training data, they won't give us scientific revolutions yet.”

Google DeepMind CEO Demis Hassabis said he thinks artificial general intelligence, or AGI, will emerge in the next five or 10 years. Of course, if you had asked him five years ago, you would have likely heard him say the exact same thing!

Google DeepMind, Anthropic, OpenAI have “talking points” on how to talk about AGI, how vague to be, how much to promise, and how much to awe (in the case of early OpenAI and Sam Altman, this included fear-mongering).

CEO of Nvidia Jensen Huang recently said that artificial general intelligence could - by some definitions - arrive in as little as five years. But what definitions are those and how many top people at his company behind closed doors would even actually agree?

AGI is Unlikely in the Near Term

If mostly machine learning researchers, by definition, don’t think it’s near - why does Silicon Valley culture perpetuate false narratives?

“ While large pre-trained systems (such as LLMs) have made impressive advancements in their reasoning capabilities, more research is needed to guarantee correctness and depth of the reasoning performed by them; such guarantees are particularly important for autonomously operating AI agents.” - Future of AI Research, Survey

It’s highly plausible or even probable that, large pre-trained models, such as LLMs, do not have the necessary adaptability, creativity, memory or complexity to realize any quantum singularity of sentience or meta-cognitive capacity. Comparing artificial systems to human capabilities isn’t just deceptive, it should probably be illegal. (OpenAI’s classic claim of its models being at or above PhD level for example is highly fraudulent)

On the topic of AGI, Chollet has expressed skepticism about Large Language Models (LLMs) being a pathway to AGI. He argues that current models are fundamentally limited in their ability to understand and generalize, particularly in out-of-distribution (OOD) tasks. Chollet suggests that the singular focus on LLMs has led to a stagnation in frontier research needed to develop AGI, as this approach risks overlooking critical advancements in other areas of AI research

In a recent paper, "Stop treating 'AGI' as the north-star goal of AI research," it’s clear that serious scientists are starting to get tired of the marketing campaign of OpenAI and BigTech here.

Hey Everyone,

My apologies for putting this out on a Saturday in the middle of the holidays, but this just could not wait.

We have an AGI update for you today. AGI is the representation of generalized human cognitive abilities in software so that, faced with an unfamiliar task, the AGI system could find a solution. We note that the commercial definition of AGI has been watered down by OpenAI, Google and many others in recent years for their systems to sound more capable. Taking this above definition however, OpenAI (employee) claims that they have AGI internally now make a bit more sense.

Share

Apparently OpenAI's o3 Model scores 87.5% on the ARC challenge (arcprize.org) - the key thing about this benchmark is that it is impossible to pre-learn, as every test has new conditions, models were stuck at 30-55%. Humans are particularly good at and LLMs were bad at it.

A bit stunning to read Francois Chollet’s blog about this. While OpenAI’s ship-mas event was fairly forgettable, the announcement of o3 models becomes a lot more interesting. OpenAI previewed their next frontier models, o3 and o3-mini, setting new benchmarks in technical capabilities and safety advancements.

While Google Gemini 2 and its announcements were fairly stunning in December, 2024 did not shape up to have any GPT-5 hype worth moments. OpenAI saw in 2024 a huge increase in weekly active users of ChatGPT and have bundled many great features into it. On December 5th, OpenAI made a new $200 a month tier called ChatGPT Pro.

But how expensive is its new o3 model going to be?

• o3 scored 87.5% on the ARC-AGI benchmark in high-compute mode, far surpassing previous records and tripling its predecessor’s performance.

• The ARC-AGI test is designed to evaluate an AI’s ability to adapt to tasks without relying on pre-trained knowledge.

• EpochAI’s Frontier Math benchmark highlighted o3’s unique reasoning capabilities, solving 25.2% of problems where others remain below 2%.

Let’s just slow down for a second, and remind ourselves that researchers view this as a major milestone but emphasize that AGI remains a distant goal, as many simple ARC-AGI tasks remain unsolved.

The o3 models use what OpenAI calls "private chain of thought," where the model pauses to examine its internal dialog and plan ahead before responding, which you might call "simulated reasoning" (SR)—a form of AI that goes beyond basic large language models (LLMs).

We all get that OpenAI needs to make more revenue in 2025 to become a viable project and have a sustainable future as a startup.

Subscribe now

Read more

Hey Everyone,

I’ve been waiting a long time for this, OpenAI has finally closed a $6.6 billion funding round with participation from the likes of Nvidia and Softbank. I believe Nvidia will be a winne…

Read more

Hello Everyone,

Today I continue my series of articles on OpenAI and the future of AI towards AGI. As you can imagine I think about this a fair bit.

As you know late last week, OpenAI released a preview of a new kind of model (Orion) chain-of-thought (CoT) models (Strawberry) branded in a preview as o1. What does this all mean?

Upgrade to Premium

(☝—💎For a limited time get a Yearly sub for just $5 a month✨— ☝).

This new CoT model for more complicated tasks is a step in the right direction for text-to-action models. However, it joins an increasing number of unfinished and not yet readily available products at OpenAI including Sora and SearchGPT, among others. Prototypes, previews, waitlists and delaying GPT-5, for a product-AI company, it’s getting fairly messy.

The o1 a is new series of AI models designed to spend more time thinking before they respond for improved reasoning. OpenAI says they are for solving hard problems but they seem to be good for mostly doing tasks a little bit better. Now I highly doubt that o1 will be as impactful as Anthropic’s Claude 3.5 Opus (Sonnet was released in June, 2024), and joins a growing list of unfinished products at OpenAI.

This new series of AI models can reason through complex tasks and solve harder problems than previous models in science, coding, and math. Although is that even true? There are a lot of moving pieces right now at OpenAI and it makes it particularly enjoyable and important for the AI landscape to analyze.

So you want to be a Product driven Company?

• Today we’ll dive into the new preview and class of models and contextualize what OpenAI is doing and must do in order to win the race to AI Supremacy.

• It’s very distracting while we wait for GPT-5 (in early 2025), to get all of these unfinished prototypes and previews of different products by OpenAI. Is this supposed to lead to AGI (Karpathy is a shareholder in OpenAI so we can’t trust his opinion here)?

For a limited time, you can subscribe to this Newsletter for as little as $5 a month with the annual plan.

Subscribe now

OpenAI has a huge team of the most expensive AI researchers in the world, and seem to constantly be trying to build shiny (new) products (Jan Leike who joined Anthropic in May, 2024). Some of their researchers behind OpenAI o1 🍓:

Microsoft Created a Monster 🧟

Read more

Hello Everyone,

Welcome to our next article in our AGI series. In it we explore what AGI means and how we might know when we get there.

It’s not always clear what Sam Altman has had in his kool-aid o…

Read more

Recently Leopold Aschenbrenner, a Columbia University graduate from 2021, revealed in a recent interview with podcaster Dwarkesh Patel that he was terminated in April for allegedly leaking confidenti…

Read more

Hello there,

This is an unusual issue of our AGI segment on AI Supremacy, for paid readers only.

Your support ensures a bright future for my independent journalism. Get a free trial to listen to the audio version of the article.

Subscribe now

Read more

Hey Everyone,

What if you could make a ‘stargate’ into the future, what would it look like? The Information recently reported on a rumored plan to build a $100 Billion AI supercomputer in the next five years. Microsoft’s project with OpenAI is raising some eyebrows.

📶 From our sponsor: 📶

Perform Deep Market Research in Seconds 📊

Using up-to-date customer insights can contribute to a 175% increase in revenue, with higher quality leads and 72% faster lead conversion… However, market research projects can easily cost thousands of dollars or more and take weeks to complete. 

Osum helps you perform deep market research in seconds. Just copy and paste a URL to instantly access detailed industry insights, SWOT analysis, buyer personas, sales prospect profiles, growth opportunities, and more for any product or business.

Stay ahead of your competition and discover new ways to unlock 10X growth.

Try Osum Now

Note: My readers gets 25% off on any plan for life. Use the code MKS at checkout.

"Stargate" AI Supercomputer is Coming

Such a project would force a radical new approach to energy, water and power consumption for the future of AI datacenters, including possibly accelerating humanity’s approach to Nuclear Fusion. U.S. President Joe Biden's administration is looking to fast-track talks with tech companies as their demands for electricity to power artificial intelligence data centers continue to grow.

A “Stargate” AI supercomputer of the scale that Microsoft might be proposing would require better energy infrastructure and could in theory far exceed current computing power for future Generative AI systems. What a Moonshot for OpenAI this would be.

They basically want to launch “Stargate” in time for the technological singularity which some believe could take place in the 2028 to 2032 window.

"AI itself isn't a problem, because AI could help to solve the problem." - Energy Secretary Jennifer Granholm

The supposed AI Supercomputer that would use millions of specialized server chips and could cost up to $100 billion, it is reported. Stargate is reportedly the fifth and final phase of Microsoft and OpenAI’s plan to build several supercomputers throughout the United States.

This is all part of America’s plan to stay ahead of China and its allies in A.I. Apparently the DOE is exploring how energy-hungry tech firms might be able to host small nuclear plants on the campuses of their massive data centers. This means BigTech is working more closely with Government than we have basically ever seen before in the U.S.

Executives at both OpenAI and Microsoft have already drawn up plans for the data center project, which would power OpenAI's artificial intelligence. OpenAI believes it is making rapid progress to what it terms AGI. They say their mission is to ensure that artificial general intelligence—AI systems that are generally smarter than humans—benefits all of humanity. But by “AGI” OpenAI seems to only mean replacing median humans at their jobs and not a sentient system that can learn in a generalized way independently.

AGI is a Moonshot for Utopia according to OpenAI

If AGI is successfully created, this technology could help us elevate humanity by increasing abundance, turbocharging the global economy, and aiding in the discovery of new scientific knowledge that changes the limits of possibility.

AGI has the potential to give everyone incredible new capabilities; we can imagine a world where all of us have access to help with almost any cognitive task, providing a great force multiplier for human ingenuity and creativity. - Sam Altman.

It’s not clear what model of OpenAI would be running on Stargate, if it is completed by 2028. Perhaps the equivalent of GPT-8. GPT-5 is believed to be the next closed-source LLM to be released.

It’s not however clear what will be the environmental, economic and energy costs of this supposed AGI? AI and data centers are going to require a lot more energy, water and resources. Since Sam Altman plans to get funding from the UAE and perhaps Saudi Arabia via a16z, it’s also not clear if such a system would be a national security risk.

Read more

Hey Everyone,

• Today of takes a closer look at where we are in terms of development of A.I. towards artificial general intelligence, or AGI.

• In recent times OpenAI, Anthropic, Google DeepMind and ev…

Read more