If AGI Has Arrived, Why Are the Builders Leaving?

The victory lap has empty chairs

Jensen Huang looked at GPT-6 Astra and posted four words on X. Maximilian Schreiner records them in Frontier Radar, his November 6, 2026 survey for THE DECODER:

“AGI has arrived.”

OpenAI president Greg Brockman welcomed humanity to the “AGI era.” The survey says the model was pre-trained with more than 100, 000 Nvidia GPUs, and that Huang said another 400, 000 were coming online. There is still no agreed definition of AGI. Superintelligence (SI, in the shorthand on some term sheets) is not a sharper target. It is a louder one.

The people who built the last several generations of frontier models are not all staying for the parade. Ilya Sutskever is gone from OpenAI. Yann LeCun left Meta in late 2025 after twelve years, following a strategy fight. Jerry Tworek, who led o1 and o3, left. Jeff Dean spent 27 years at Google, most recently as chief scientist, and left in early August 2026 with Sanjay Ghemawat, Quoc Le, and Oriol Vinyals, as the survey reports it. Mira Murati and John Schulman are building elsewhere. The exits became a visible wave in 2024. They have not reversed.

That is not proof that powerful AI is far off. In a late 2025 conversation with Dwarkesh Patel, Sutskever still put a system that learns as well as a human somewhere between five and twenty years out. Tworek left to work on learning after deployment, which is a strange errand if you think the recipe is finished. LeCun left to back a different architecture, not to close the lab. Richard Socher, who is not on that exit list, said the definitional part out loud: “Depending on your definition, we’re getting close to AGI.”

The doubt that should touch a budget is narrower than “AI stops here.” The current recipe (ever-larger language models, trained on internet data, then polished with reinforcement learning) will not be enough by itself.

A bigger oven is not a new cuisine

The oven is still getting bigger. Huang’s extra GPUs are a purchase order, not a metaphor. What they bake is the same cuisine: more parameters, more scraped text, more reinforcement learning on top.

In the usual setup, a deployed model does not update its weights. It can take notes, retrieve documents, and call tools. Those are memory aids. Continual learning is something else: the model changes its weights from experience after it ships, and a new skill sticks. Stuffing a fact into the chat window is not that.

Retrain carelessly and you get catastrophic forgetting. The network writes over what it already knew. Leave the weights frozen, and your AI agents stay brilliant interns who forget the company every night. Useful for a draft. Useless as institutional memory.

Sutskever’s five-to-twenty-year range is an admission that the calendar is still a guess. You cannot treat it as a product roadmap. Tworek has put the career version more bluntly. In a race for the best model, you take fewer risks. The giants are in that race. The people walking out are trying to fund the risks the race punishes.

Three tests, before the tour

If you run a P&L, hold three tests in your head. The labs are illustrations. They are not a shopping list.

Keep shipping AI agents and AI automation on models you can buy this quarter. Notes, retrieval, and tool use are good enough for a lot of work, if you measure error rate and unit cost instead of applause.

Do not confuse that stack with a memory. A model that cannot update its weights will not become the system that knows how your company actually works. If continual learning ships, the retrieval wrappers you standardized on start to look like filing cabinets. Budget for a swap. Do not pretend the apprentice is already on payroll.

Read a product-less lab as a small, reversible bet, not the platform you standardize on. Ask one question of every pitch. Are you paying for a feature a frontier lab can copy in a news cycle, or for data, a fleet, or an audit trail?

Why investors fund a lab with no demo

Schreiner’s word for the new shops is unofficial. A neolab is an independently founded, research-driven company built around a long technical bet: how AI learns, perceives, acts, stays under control, or computes. As of that November 6, 2026 survey, he says he sorted 45 of them, mostly in the United States and Europe, with 27 profiled in detail. It is one analyst’s map, not a census.

The investor logic is old. The valuations are not. OpenAI began as a nonprofit in 2015. Anthropic was founded in 2021 by Dario and Daniela Amodei after they left. Those two are now large companies, widely described as heading toward public markets at enormous valuations. By the end of July 2026, Bloomberg had put Anthropic’s annualized revenue above $65 billion. A run-rate is not cash in the door. It is enough to change what a research meeting feels like. Product pressure and a public-market story crowd out ideas that threaten the model you already sell. That is the innovator’s dilemma with a term sheet attached.

The precedent should bother any incumbent. Google invented the Transformer in 2017. DeepMind built AlphaGo. Neither shipped the mass-market assistant. A younger independent lab did. A spinoff of that lab became the other giant. Breakthroughs do not have to cash out at the inventor’s company.

Do not romanticize every resignation. Some founders want a technical argument the giant will not fund. Some want paper wealth. Some are tired of seeing their niche treated as a sideshow inside a chatbot company. And some breaks are not breaks. Google is a founding investor and cloud partner in Discovery Loop, the lab Dean and his colleagues started. WIRED reported the idea was only weeks old when they left in early August 2026.

Two checks you can actually cite

Most money headlines in this wave still live in a single survey, a secondary-market mark, or the word “talks.” Two of them have a news report underneath. Even those need careful verbs.

On July 27, 2026, Bloomberg reported that Nvidia has committed to invest $5 billion in Sutskever’s Safe Superintelligence, according to people familiar with the matter. Bloomberg called it one of Nvidia’s largest funding deals of the boom. The joint statement that day did not disclose terms. What it put in public was access to Nvidia’s next-generation Vera Rubin platform. In that announcement, as the survey renders his line, Sutskever said SSI had research that deserved to be scaled up. As of the November survey, there is still no public model to inspect. Treat the $5 billion as a reported commitment, not as a number the companies printed in a press release, and not as a completed wire.

Prometheus is louder about the round and quieter about the work. On June 11, 2026, CNBC reported that the company raised $12 billion in Series B funding at a valuation of roughly $41 billion. Jeff Bezos, who stepped down as Amazon CEO in 2021 and remains executive chair, is co-CEO with Vik Bajaj, a Stanford School of Medicine professor and co-founder of Alphabet’s Verily. Bezos has said Prometheus is the bulk of his time, alongside substantial time on Blue Origin and on AI at Amazon. He dated the work to late 2024. He said he joined the Series B after putting capital into a first round. That earlier amount has not been disclosed.

He calls the target an “artificial general engineer”: AI for invention and physical engineering, including manufacturing and, in CNBC’s account, drug design. He has denied that this is a robot company. He has also said it is still “premature” to disclose accomplishments. CNBC says the company has shared few product details. Secrecy is not proof that nothing exists. Bezos put the team at about 150 people, recruiting from OpenAI, Google DeepMind, and Nvidia.

The line that matters for every other neolab is why the round is so large. A big share of the money is for compute. The work is compute-hungry because the data has to be created. Internet text will not design the part that does not exist yet. Compute is scarce enough, he said, that they take it from many places, including AWS. He called it “easy to imagine” Prometheus as an Amazon cloud customer, and hyperscalers using Prometheus tools, while the companies stay “at arm’s length.” Easy to imagine is not a signed partnership.

Six bets against the recipe

Schreiner groups the field into six bets. Read them as a map of disagreement. Unless a news outlet is named, funding figures below are survey claims as of November 6, 2026. Secondary-market marks, including numbers that circulate under the Forge name, are chatter until a round closes. Most of those marks are left out on purpose.

Learning after the model ships

The disagreement: a model that cannot change its weights after deployment may never be more than a brilliant intern with amnesia.

Sara Hooker, now at Adaption with Sudip Roy (Adaption is the company name, spelling and all), has argued since at least her 2020 essay “The Hardware Lottery” that methods win partly because they fit the chips already on the shelf. The survey also points to her later claim that extra compute is improving models less reliably than the scaling story promised. Adaption is selling data tools and an AutoScientist API. The continual-learning model is still missing. The survey lists a $50 million seed in February 2026.

Core Automation is the sharper version of the same bet. Tworek’s co-founder, Rohan Anil, led Gemini pre-training. The rumor, and it is still a rumor, is a first model that updates weights in deployment and trains on about a hundredth of today’s data. No public product has been confirmed.

François Chollet’s distinction is the buying test, whether or not you ever meet his lab. Solving a task the model was trained on is ability. Learning is acquiring a new skill cheaply. A leaderboard often sells you the first and invoices you for the second. If Core Automation, Adaption, or someone quieter actually ships weight updates that stick, the systems you wired around retrieval are the filing cabinets. Until then, do not pay a platform price for a rumor.

Worlds, space, and robots

The disagreement: predicting the next word may be the wrong target, and intelligence may have to model space, physics, and action.

LeCun’s AMI is the prestige version. JEPA, which he proposed in 2022, predicts abstract representations instead of pixels or words. At Meta, V-JEPA 2 produced a narrow 2025 result the survey highlights: after video pre-training, 62 hours of robot data was enough for simple pick-and-place in unfamiliar labs. AMI raised $1.03 billion in March 2026 at a $3.5 billion pre-money valuation, the survey says, with about a dozen people. CEO Alexandre LeBrun has said a product could be years away. Research and most of the code are planned to be open. That is a research institute with a venture cap table.

Fei-Fei Li’s World Labs is further along commercially, if the survey’s timeline holds. Marble has been selling since November 12, 2025, with an API since January 2026. Atlas, for select partners since September 2026, turns a few photos into a walkable space and up to a minute of video. Her line for the bet, relayed via Possible, is that 3D is the language of nature. AMD has agreed to acquire World Labs for about $8.2 billion in stock, with a close expected by the end of 2026 and Li slated to become AMD’s chief scientist. Pending is not closed. If it closes, AMD is buying a spatial-intelligence story rather than waiting to grow one beside Nvidia’s.

The invoice is the part a plant manager should read. On an AA-Robotics comparison the survey cites, Astra leads with 69 of 100 points, at more than $16 a task, against a human baseline of about 20 seconds. Leading the scoreboard while losing the clock and the cost sheet is not a deployment plan. General models can look competent at spatial work by orchestrating simulators, planners, and control software that already existed. Competence at that price is a demo.

Data you cannot scrape

The disagreement: the scarce asset may be measurements nobody posted on the internet, rather than a cleverer training trick.

Methods leak. A private dataset, a robot fleet, or an automated lab is harder to photocopy. Bezos already said the data has to be created. Periodic Labs is that sentence with a lab attached. Its motto is the cleanest line in the category:

“Intelligence is necessary, but not sufficient.”

Neon, unveiled September 15, 2026, is an existing trillion-parameter model fine-tuned on literature and lab data, the survey reports. The company claims it beats GPT-6 Astra and Claude Fable 5.1 on X-ray analysis at lower cost. The first named customer is a chipmaker with a heat problem. Public model access has not been confirmed. A company claim is not an independent benchmark.

Basecamp Research, in London since 2019, says it samples with partners in more than 30 countries and holds more than 100 billion previously unknown genes. The Trillion Gene Atlas, announced in March 2026 with Nvidia and Anthropic, is a data moat dressed as a science project. Six cell-therapy programs sit in early research, with no clinical-trial timeline in the survey. None of this is a chatbot feature. The argument is that AI for business, in the physical economy, starves without measurements the internet never had.

Keep the cautionary twin on the same page. GNoME, the 2023 DeepMind system, predicted millions of crystal structures. A few hundred have been confirmed in the lab, the survey notes. Prediction is cheap. Confirmation is the business.

Compute that is not just more GPUs

The disagreement: the next constraint may be how an answer is produced, and what that production does to the power bill, rather than parameter count.

Inception’s Mercury 3 shipped on October 19, 2026. The company says it beats smaller GPT-6 Luna and Claude Haiku 5.5 on eight of twelve benchmarks and runs four to eight times faster, by drafting a whole answer and refining many parts at once rather than speaking one token at a time. A Transformer can still sit underneath. Speed is a product. It is not a new theory of mind. Reuters has reported Microsoft acquisition talks here. More on that in a moment. Talks are not a product either.

Liquid AI, founded in 2023 by MIT researchers, is the one already in a car. The Wall Street Journal reported on November 5, 2026 that Mercedes had begun rolling the technology out in North America in October: simple voice commands on the device, harder requests in the cloud. That hybrid split is what most AI automation programs will actually look like. Not one genius model. A cheap reflex and an expensive brain.

A cluster of labs wants to leave ordinary transistor switching behind, using oscillators or thermodynamic hardware. Naveen Rao expects energy to become a hard limit in about three years. The startling efficiency numbers in this lane mostly come from simulations. File them under a power bill you might face later, not under a stack rewrite this quarter.

Machines that do the research

The disagreement: the fastest path to stronger AI may be a system that runs the research, rather than a better base model.

The survey sketches four stages. Assistance. Then an automated trial inside a fixed search. Then open-ended work on a problem a human still chose. Then a system that builds its own successor. Almost everything public sits at stage two. A few projects touch stage three in a narrow lane.

Axiom, founded in 2025, is the exhibit that will either impress a research chief or sink the slide. The company says dozens of agents, under human guidance, formalized the Enormous Theorem (the classification of finite simple groups) in the Lean proof language in 24 days. The underlying mathematics runs past 10, 000 pages, across hundreds of papers and five decades. Axiom contrasts its run with one section of about 250 pages that had taken a human team six years in an earlier system, Coq, and points to generated code on the order of 43 million lines. The survey also says the effort surfaced an error in a 1986 paper. It relays mathematician Timothy Chow saying experts knew it, but that it had apparently never been written down in public. Treat all of that as the company’s showcase until you, or someone paid to be skeptical, have read the repository. Lean certifies the formalization you wrote. Translate the question badly and you prove the wrong statement with a perfect checkmark. The survey lists a $200 million Series A, with a valuation reported from $1.6 billion.

The incumbents are not watching from the stands. OpenAI said in September 2026 that it had reached an “automated research intern, ” and that it wants a fully automated AI researcher by March 2028. Anthropic reported in June 2026 that Claude writes more than 80 percent of newly integrated production code. Use that figure, not the rounder numbers that float around it. The survey also notes an OpenAI proof for a variant of the Navier-Stokes problem, published in September 2026 and not yet accepted, with more than 700 manuscripts following in October. Automation produces papers. It also produces sludge.

If agents write most of the new production code, permissions, logs, and a kill switch are the safety case. The stage label is not. A sandbox nobody is required to use is not a sandbox.

The gaps around models you already rent

The disagreement: you may need the model you already rent to be honest, calibrated, and inspectable. A new frontier model is a separate question.

Goodfire’s RLFR work, reported in February 2026, put hallucinations down 58 percent on a 12-billion-parameter Gemma 3, at about 90 times lower cost per intervention than an external judge. Read the measurement before you buy the slide. It includes monitoring, correcting answers in deployment, and picking the best of 32 responses. The later step, mechanistic shaping, published November 5, 2026, penalizes the model during training for leaning on misleading internal signals. That is a different claim from “the model stopped lying.” Partners listed in the survey include Microsoft, Mayo Clinic, and the Arc Institute. For regulated work, inspecting the machinery beats trusting the model’s own explanation of itself.

Decision models are the sleeper product for AI for sales and for any approval workflow. They emit a choice from a fixed menu, plus a confidence score. Calibration is the whole trick. If the model says 80 percent, it should be right about 80 percent of the time in similar cases. TypeSafe launched Jev in mid-September 2026. OpenAI announced a Decisions API two weeks later, generally available since October 22. Cloudflare has Clef. The survey puts TypeSafe’s Series A at $870 million on October 9, 2026, at a $7.5 billion valuation. A two-week head start became a fundraising event. It did not become a moat.

A two-week lead is not a moat

Jev, then the Decisions API, is the pattern in miniature. A small lab’s method can show up, renamed, at a market leader before the next board packet. Even a short lead can end as an acqui-hire: the buyer wants the team more than a durable product. Microsoft hired Inflection’s core team. Google hired the Character.ai founders and Windsurf’s leadership. Reuters has reported Microsoft acquisition talks with Inception. Inception wanted more than a billion dollars. The Financial Times has reported Nvidia talks with Reflection. Nothing in those last two is confirmed. Talks are not a close.

Incumbents also have the dull advantages. More compute. Thicker data pipes. Feedback from hundreds of millions of users. A prompt trick is not a moat. Exclusive data and physical infrastructure are the exceptions worth underwriting. A dataset you cannot scrape, a robot fleet, a lab that generates its own measurements, a formal checker with an audit trail: those are harder to photocopy over a weekend. Ask which one is on the invoice.

The survey counts Nvidia as an investor in at least 25 of the 45 labs, with technology or research ties to four more, across every bet. The reading that follows is hard to wave off while it lasts. Contradictory scientific bets still tend to rent the same chips, so capital Nvidia puts in can come back as compute spend. The same logic gets applied to Google, Amazon, and AMD. That is a position, not a law. The two deals with public reporting behind them (a Vera Rubin access arrangement with a reported $5 billion commitment, and a Prometheus round that is openly a compute purchase) show the pattern. They do not prove that chipmakers win if the science leaves GPUs behind. Analog hardware, serious on-device models, or a real energy wall could move who captures the spend. Until one of those shows up in a purchase order, the documented pattern is that quiet labs raise money in order to buy compute.

Two or three years, not a prophecy

Schreiner’s baseline for the next two to three years should humble a strategy offsite. Funding stays high. Consolidation follows. Few methods stay exclusive. Many teams get bought, acqui-hired, or marked down. Acceleration, in his judgment, is plausible only where a lab can show results a third party can check. Materials research is the example he allows. For some product-less labs, a markdown or a soft landing inside a giant is the expected outcome. That would not, in his view, break the wider dynamic. It would be the dynamic.

Discovery Loop is the froth in one paragraph. The first round is confirmed. The amount is not. Reported talks moved from about a $1 billion seed at a $10 billion valuation toward figures around $50 billion. Those are negotiation numbers. There is no public product. If you are modeling superintelligence by the next board meeting, you are modeling a slide.

Write it into the budget

Do not freeze an AI automation program because a founder declared an era. The current recipe already runs the agents, copilots, and sales assistants you can buy this quarter. Ship them. Instrument them. Kill the ones that fail a cost test.

Do not sign a multi-year exclusive with a lab that cannot show you a system. Assume the team may sit inside Microsoft, Google, Nvidia, or AMD before the renewal. Write the exit into the contract: data export, a model-swap test, and a definition of “done” that does not depend on the founder still being independent.

For AI for sales, ignore the banner and ask whether confidence is calibrated. A decision model that says 70 percent, and is right about 70 percent of the time, is more useful than a chatbot that is confident about a fantasy pipeline. For anything that touches money, patients, or a factory, ask who can inspect the failure. A 58 percent improvement that includes “we picked the best of 32 answers” is a process change. It is not a cured model.

And treat agents that write production code as an access policy, not a research aesthetic. If an agent can merge code or call external tools, the permission boundary comes before the vision deck. Optional sandboxes do not count.

The replacement budget is the part most plans skip. Continual learning, private scientific data, and a decision layer with honest probabilities are not in the building just because the term sheet exists. Leave room for them anyway. When one of them becomes real, the workaround you standardized on will look exactly like what it is: a clever filing system next to an apprentice.

Questions worth asking before you rewrite the roadmap

• If AGI has arrived, why are the builders leaving?

  “Arrived” is a marketing sentence, and there is still no agreed definition. The departures are a bet that the path is more open than the giants’ messaging, not a verdict that stronger AI is far off.

• What is a neolab, in practice?

  An independent, research-heavy company built around one long technical bet, often with little or no public product. Schreiner’s November 2026 survey says it sorted 45 of them. You are funding an argument, not a category you can standardize on.

• Is scaling dead?

  No. The live argument is that more of the same improves models less reliably, and that data and learning methods matter more. More compute still does something. It may not do the thing the victory lap is selling.

• Which money figures are firm?

  Two, and only with careful verbs. CNBC reported Prometheus’s $12 billion Series B at roughly a $41 billion valuation. Bloomberg reported a $5 billion Nvidia commitment to SSI via people familiar, while the joint statement disclosed Vera Rubin access and not the price. Read the rest as survey claims, secondary marks, or talks.

• Will Nvidia win no matter what?

  That is the survey’s lean, and it is plausible while almost every bet still buys GPUs. The SSI and Prometheus deals fit the pattern. It is not a law: analog compute, on-device models, or an energy wall could move who captures the spend.

• Am I paying for a feature a frontier lab can copy, or for data, a fleet, or an audit trail?

  If a giant cloned it in two weeks, you bought a feature. A dataset you cannot scrape, a robot fleet, a lab that generates its own measurements, or a checker with a real audit trail is a different invoice. Contract for the thing you can retest.

• What should change in an AI budget this year?

  Keep shipping agents you can test on cost and error. Do not sign a multi-year exclusive with a product-less lab. Leave room for continual learning, private scientific data, or honest probabilities to make part of today’s stack look like a workaround.