Neolabs は超知能への賭けか、AGI 実現の現実的課題を問う

I have been forecasting frontier lab progress for years now. My team was the first to figure out an accurate breakdown of OpenAI's revenue, I called Anthropic's rise to the top lab of 2026 back in January, having tracked its financials on the way up, and we co-authored the AGI timelines behind AI 2027.

So I am increasingly puzzled by people putting billions into these new startups going after AGI. How do they hope to compete, if even Google and Meta and xAI can't keep up?

The standard answer is that these "neolabs", built around famous researchers and enormous checks, have differentiated approaches. Reflection AI is open source, LeCun's AMI is betting against LLMs, Thinking Machines is commercializing B2B early. (David Silver's Ineffable, and Ilya's SSI, are more mysterious.)

This doesn't really explain Discovery Loop. Anthropic and OpenAI are both on the record for trying to automate R&D. It's a core part of the AI 2027 timeline forecast that FutureSearch co-authored, the one that led us to predicting superhuman capabilities around 2031.

My conclusion is that people betting on neolabs do not believe in recursive self-improvement. They are betting against superintelligence, and they think LLMs will plateau, despite every such prediction so far being spectacularly wrong.

It's a funny position for a VC. "Sure, I'll put $1B into this exciting AI startup, founded by this famous AI researcher. But no, I don't expect AI to take off anytime soon." It's not illogical, it's just a narrow space of possible futures: AI is massively disruptive, but only in 10+ years from now, and possibly with different approaches to what we're doing now.

So I forecast all six of the "AGI" companies that have multi-billion-dollar war chests: Safe Superintelligence, Thinking Machines Lab, Reflection AI, David Silver's Ineffable Intelligence, Yann LeCun's AMI Labs, and Discovery Loop (founded this morning). Each name links to a dedicated forecast page I have since written on that lab, carrying the questions this comparative table leaves out: release dates, revenue, and valuations.

This is a bit hard to forecast, which you can see in the very wide confidence intervals. The most interesting lab on the list ships nothing on purpose, so any question that resolves on products or leaderboards fails on SSI. Here, a frontier model means top-5 on a recognized index if the lab releases models, or credibly established as frontier-class by reporting and independent expert assessment if it does not.

*Three labs racing for the frontier around 2029 to 2030, and three research bets

whose medians sit a decade (!) out.*

Unsurprisingly, the capital, compute, and talent at these neolabs is a tiny fraction of OpenAI and Anthropic. The best compute medians on the bench, three-quarters of a gigawatt, sit an order of magnitude below the ten-gigawatt scale OpenAI's Stargate buildout is heading toward, and the best capital medians are about a sixth of OpenAI's latest round alone. If transformative intelligence arrives on the incumbents' timelines, it arrives from the incumbents. The neolabs' chances of being competitive for AGI, the way the race is currently going, are obviously low.

What follows are my summaries of each lab, drawn only from the rationales behind the four questions I asked about all six of them, and weighted toward the timeline I care about most, when each one first has a frontier model. These are not full profiles. Every lab has its own forecast page, linked from its name above, carrying the release dates, revenue, and valuations this bench leaves out.

Safe Superintelligence

SSI is the most credible challenger on every resource gate and the least observable on every output gate, as per the forecasts. It has quietly raised about $8 billion, more than double the figure most coverage carries, counting Nvidia's $5 billion July investment on top of roughly $3 billion across its earlier rounds. The Nvidia deal comes with priority access to the next-generation Vera Rubin platform, which the companies say raises SSI's compute by an order of magnitude. That buys the earliest frontier median on the bench, January 2029, with a left tail reaching March 2027.

The surprise is the talent number. Only two people at SSI are publicly verifiable as former senior staff of the big three labs, according to FutureSearch, and the median forecast a year out is just 7, on a total headcount around 50. Every other serious lab treats hiring volume as the weapon. SSI is betting that capital and compute concentrated on the smallest possible team beats headcount. (Remember when OpenAI had a hard limit of 150 employees, and even let people go to enforce that?) What a distant observer should watch for is not a launch, because there will not be one. It is the tone of the reporting, since the frontier question here resolves the way Anthropic's unreleased Mythos did, through leaks and independent expert assessment.

Thinking Machines

Thinking Machines has the largest compute anchor on the bench, an Nvidia deal for at least a gigawatt of next-generation systems beginning early 2027 plus a single-digit-billions Google Cloud expansion for reinforcement learning, and it has actually shipped a foundation model, which really distinguishes it from this list (even though the model isn't very good). It's called Inkling, a 975-billion-parameter mixture-of-experts, debuted at 41 on the Artificial Analysis index, thirteenth of roughly a hundred models tracked and far from the top five that we're forecasting about here.

That is why the lab with the best infrastructure carries a frontier median of December 2030 and a right tail to 2043. The product-first path, Tinker and mid-tier models, gives it something to sell while the frontier bet waits, possibly indefinitely. Thinking Machines is the lab most likely to become a great business without ever holding the frontier.

Reflection AI

Reflection is the sleeper, and the FutureSearch forecasters found some things I hadn't seen in the news. It has raised about $4.6 billion, with a March Series C at a $25 billion pre-money that got surprisingly little coverage, and it is the fastest-hiring lab on the bench, about 230 people today, up from roughly 60 last fall, with a median of 32 senior big-three hires by next August, the most on the bench and half again Thinking Machines' 23. Its open-weight strategy also gives it the most visible path to resolution, which is why its frontier median of June 2029 sits within striking distance of SSI's despite far less capital.

Reflection doesn't have much compute. (Though this never stopped Anthropic.) Reflection's footprint is rented month to month, $150 million per month on SpaceX's Colossus 2 campus under a lease either side can end on 90 days' notice. It is the structural opposite of SSI, people-heavy and infrastructure-light, and the thing to watch is whether it converts talent into a flagship model before the rolling lease or the burn rate forces a choice. Its flagship open-weight model has yet to ship.

Ineffable Intelligence

David Silver's lab holds Europe's largest-ever seed at $1.1 billion, and the surprising line in its cap table is the UK's Sovereign AI fund sitting alongside Sequoia, Nvidia, and Google, a national government taking a direct position in a pre-product research lab. They could have good infrastructure, with one of the largest next-generation Nvidia clusters on Google Cloud, announced in June.

And Silver is not sitting out the AGI race. (I'm reading a bit about him in the new Demis Hassabis book.) He is betting a different architecture wins it, calling human data a fossil fuel and building systems that learn from experience instead, which I take has been his view since the early DeepMind days, it's a very AlphaZero perspective. The late frontier median, June 2035 with a tail past 2050, is not a verdict on the ambition. It measures how long before a lab that rejects the LLM paradigm fields something leaderboards can see. Watch for reinforcement-learning results.

Discovery Loop

The most famous founding team on the bench has the second-lowest compute median and a frontier median of June 2036, which is wild to me, will anything other than frontier labs matter by then? The forecasts read Discovery Loop as a consumer of frontier models rather than a producer, an orchestration layer for automated research riding on Google as founding investor and cloud partner. They will still probably want models of their own eventually, since a loop that automates ML research is also a loop that trains models, and the August 2029 left tail carries exactly that scenario. The median just says the route is long, and the 2060 right tail says it may never arrive at a flagship model at all.

So the question is: even if they do automate R&D, can they use their automated R&D or will they just sell it to someone else who will build AGI and eat the world economy?

I published a full set of Discovery Loop forecasts in a dedicated piece, including which scientific field its loop would crack first. The lesson it adds to this table is that the neolab category conflates two different kinds of company. Is this even an AGI company?

AMI Labs

Yann LeCun's lab is last on every metric, a sixteenth of SSI's compute median, the smallest capital trajectory after its $1.03 billion March raise, six senior hires, and a frontier median of December 2037 with the longest tail on the board. Apparently this is intention. LeCun has said human-level AI is not going to be built on LLMs, and the lab's world-model architectures are invisible to a text-only leaderboard, so we're measuring a company running a different route than the one it describes. LeCun's claim is that LLMs won't reach AGI, so for him maybe a decade long project actually makes sense?

If AMI is right, its success will barely show up on these metrics, and the first sign will come from somewhere none of these questions look. For a bench built to measure the race, one lab that rejects the premise is a control group worth having.

Grading the forecasts

I hope we get more evidence on the neolabs soon. I guess my main prediction, having reviewed all of this is: by the time any of them have anything serious, either OpenAI or Anthropic will have built superintelligence.

Forecast these yourself in the

FutureSearch app

, the moment DeepMind ships, or the moment someone else leaves.