The recipe is exhausted; the curve is not

The most-misread sentence in the scaling debate dates from December 2024, and the words that get dropped in the misreading are doing all the work. At NeurIPS 2024 in Vancouver, accepting the conference’s Test of Time award for the 2014 sequence-to-sequence paper he co-wrote with Oriol Vinyals and Quoc Le, Ilya Sutskever — who co-founded OpenAI, served as its chief scientist for more than eight years, departed in May 2024 and now runs Safe Superintelligence — said: “pretraining as we know it” will end. The headlines that ran the next day, and the secondary pieces that ran the week after, said Sutskever had called time on scaling. He hadn’t. He’d said the recipe was finished. Sutskever did not say pretraining ends; he said the version of it that runs on the public internet does.

The distinction matters because the next sentences out of Sutskever’s mouth in Vancouver were not bearish on compute. He called data “the fossil fuel of AI.” He said “we have but one internet.” Both phrases mean the same specific thing: the supply of high-quality natural training text that pretrained GPT-3 and GPT-4 is finite, has been substantially exhausted at frontier scale, and cannot be re-created. Compute, meanwhile, keeps growing. The architecture keeps growing. What ran out is the cheapest input. A scaling-is-over reading collapses these into a curve claim. A recipe reading does not.

Eleven months after Vancouver, on November 25, 2025, Sutskever clarified in a 90-minute interview with the podcaster Dwarkesh Patel, his first long-form since leaving OpenAI and his second with Patel overall. He framed 2012–2020 as the “age of research,” 2020–2025 as the “age of scaling,” and 2026 onward as a renewed “age of research” carried out, this time, on much bigger computers. SSI, he said, is “squarely an age of research company.” Then, rejecting a Sam Altman framing he’d been asked about, he asked his own question of the bull camp: “is the belief that if you just 100x the scale, everything would be transformed? I don’t think that’s true.” He did not say more compute would do nothing. He said it would not be transformative under the current recipe. The bull camp has misread the qualifier back at him almost as consistently as the press did.

What the qualifier actually does

Without the qualifier, Sutskever is saying scaling is over — and the industry that has just committed something close to two trillion dollars to compute infrastructure between OpenAI and Anthropic alone has therefore made the largest capital-allocation error in the technology sector’s history, and the man saying so is operating a $32 billion lab on Google Cloud’s TPUs. The reading is internally incoherent.

With the qualifier, Sutskever is saying the specific recipe of web-scale autoregressive pretraining cannot be extended further on data alone. The other axes — reinforcement learning post-training, inference-time compute, verified-reward environments — are scaling, and Sutskever does not contest that they are scaling. He says, on Patel’s podcast, that practitioners on Twitter report spending more compute on RL than on pretraining now, characterizes RL as delivering “a relatively small amount of learning per rollout” relative to compute consumed, and notes that “we’ve already witnessed a transition from one type of scaling to a different type of scaling.” That reading is consistent with running SSI on TPUs, with raising $3 billion at $32 billion to do whatever it is he is doing on those TPUs, and with a public statement calendar — one X post through the spring of 2026, on Anthropic’s Pentagon designation, not on technical questions, followed by the NVIDIA announcement in July and, on September 1, a warning that neoclouds had limited cybersecurity and that labs with strong cyber models should help defend them — that suggests deliberate institutional alignment with the labs that are still scaling.

What he is not contesting

The strongest evidence for the recipe-not-curve reading is the body of cost-and-capability data Sutskever has had every opportunity to engage with publicly and has not. The Gundlach et al. measurement in “The Price of Progress” (NeurIPS 2025 Workshop; revised March 2026) is that the price of any fixed level of frontier benchmark performance has fallen roughly five- to tenfold per year on the benchmarks his recipe-exhaustion claim does not touch.

Gundlach et al. add a warning the bull camp also does not foreground: while the cost of holding any given capability has collapsed, the cost of holding the frontier — the cost of being on the leaderboard at any given moment — is rising. Demand for larger models and more reasoning tokens is consuming, and at the frontier overcompensating for, the per-token efficiency gains. That is consistent with both readings of Sutskever. The recipe reading says: the falling price curve is the second axis turning over, exactly as Sutskever’s own framing predicts, while frontier-record cost rises because the new axes are expensive. The curve reading would say: nothing has bent.

What Sutskever has not done is contest any of the underlying cost-and-capability data. He has had a publicly available primary source — the Patel transcript, the SSI launch text, the NeurIPS recording — to dispute the empirical record, and he hasn’t. The cost-per-capability frontier through November 2025 is not what his thesis needs to deny.

The institutional bet, placed before the technical argument

SSI announced its founding on June 19, 2024, six months before the NeurIPS talk. The first $1 billion raise closed September 4, 2024, three months before. The Greenoaks-led round, which valued SSI at $32 billion, closed in April 2025, four months after. SSI moved to Google Cloud TPUs as its primary compute on April 9, 2025. By July 3, 2025, after Daniel Gross departed for Meta, Sutskever was CEO, and he wrote to SSI’s team and investors: “We have the compute, we have the team, and we know what to do.” None of that infrastructure is consistent with curve skepticism. All of it is consistent with a belief that a different recipe, executed on the same substrate the bull labs are using, will get to a different place.

The contrast with the public registry of Sutskever’s words is sharp. Two podcast appearances in eighteen months. A handful of X posts. One conference talk. One January 2026 NAS Award — the first time the National Academy’s Industrial Application of Science award has gone to AI, announced while he was making no public statements at all. The institutional record is not thin. The institutional record is a multi-billion-dollar lab running at frontier scale and conducting research the public cannot see.

In October 2025, MIT Technology Review published a piece characterizing Sutskever as the central figure in an “AGI as conspiracy theory” frame; The Atlantic in 2023 reported that he had led “feel the AGI” chants at internal OpenAI events. The aura is part of why writers reproduce his categories without examining whether they survive the institutional bet he is placing alongside them. They mostly do. The categories survive only if they are read as recipe categories, not curve categories.

Who disagrees, and how

Dario Amodei, Anthropic’s co-founder and CEO, is the cleanest counter. On February 13, 2026, also on Patel’s podcast, Amodei said: “my top level model of scaling is the same as it was in 2017.” On March 3, 2026, at Morgan Stanley’s TMT conference, he said: “We do not see hitting the wall. We don’t see a wall.” His “country of geniuses in a datacenter” scenario, the framing he coined in Machines of Loving Grace in October 2024, sits at 90% likelihood on a ten-year horizon. The two of them are using different instruments. Sutskever is asking whether the recipe is exhausted; Amodei is reporting the curve has not bent. The Amodei position concedes nothing structural.

Altman is the noisier counter and, given OpenAI’s $1.4 trillion in announced infrastructure commitments, the louder one. His February 9, 2025 “Three Observations” post argued that “the cost to use a given level of AI falls about 10× every 12 months,” that “the intelligence of an AI model roughly equals the log of the resources used to train and run it,” and that “the scaling laws that predict this are accurate over many orders of magnitude.” The Gundlach et al. measurement broadly validates the cost-per-capability framing without contradicting Sutskever’s recipe-exhaustion claim, because the Gundlach paper measures inference prices at fixed benchmark performance and Sutskever’s claim is about the data input. The two arguments slide past one another. That is part of why the misreading sticks: the bull camp’s strongest empirical claim and the recipe-rejection claim are about different objects.

Noam Brown, OpenAI’s vice president of research, whose work powered the o-series move into inference-time compute, has been more pointed. His April 23, 2026 post on X: “with today’s AI models, intelligence is a function of inference compute. Comparing models by a single number hasn’t made sense since 2024. What matters is intelligence per token or per $.” Brown’s claim is that the curve has acquired a second axis — inference-time compute — that pretraining-only readings cannot see. The argument is unfinished; whether the second-axis reframe is the bull case’s strongest move or its tell that the original axis has stalled is not yet resolved. For now, the move is real, the data behind it (smooth log-linear improvement on reasoning benchmarks with thinking-token budget, replicated across OpenAI’s o-series, Anthropic’s Extended Thinking, and Google DeepMind’s Deep Think) is real, and it is — by Brown’s own framing — the central argument the bull case is now making.

The piece of Brown’s argument Sutskever has come closest to engaging is the cost claim. On Patel’s podcast he characterized RL as delivering “a relatively small amount of learning per rollout” relative to compute consumed. That is not a denial that inference-time compute is a scaling axis. It is a denial that the slope is what the bull camp says it is.

What the next twelve months actually test

A recipe claim and a curve claim are different falsifiable claims. Sutskever’s recipe claim — pretraining-as-we-knew-it cannot be extended on data alone — is mostly settled at the frontier; the bull labs have largely conceded that the next gains come from RL post-training, from inference-time compute, and from verified-reward environments rather than from scraping more of the public internet. What is not settled is whether those new axes, operated together, keep delivering the cost-per-capability gains that Gundlach et al. measured through 2025. That is the test for the curve claim.

Two tests bear on it. ARC-AGI-3, the interactive benchmark the ARC Prize Foundation launched on March 25, 2026, opened with the December 2025 frontier models below 1% against 100% for humans. On September 2, 2026, ARC Prize reported GPT-6 Astra at 62.7% on the Semi-Private set under its Standard harness at max reasoning effort. A benchmark built to test how fast a system learns unfamiliar games went from near zero to a majority score in under six months, and the architectural critique adjacent to Sutskever’s has to absorb that. If Gundlach et al.’s measured rise in frontier-record cost accelerates, Altman’s tenfold-a-year framing has to be footnoted as holding at fixed capability rather than at frontier capability. Neither test touches Sutskever’s recipe claim. Both touch the curve.

Sutskever himself has said almost nothing about which way these tests will go. SSI is not shipping. The “missing machine learning principle” he hinted at in Vancouver and elaborated on aesthetic-criteria grounds — “beauty, simplicity, elegance, correct inspiration from the brain” — in November 2025 has not been disclosed. The public bet is the institutional bet: a $32 billion, fifty-person lab, in Palo Alto and Tel Aviv, running on Google’s TPUs, betting that a recipe yet to be written, executed on the substrate the bull labs are also betting on, gets somewhere worth getting. The other half of his bet — that what comes next is qualitatively different from what came before, agentic and reasoning and unpredictable in the AlphaZero sense — is the part Brown’s framing converges with rather than contradicts.

The argument that started in Vancouver is, on a careful reading, not a fight about the curve. It is a fight about which recipe is on the curve. Sutskever’s view is that the next recipe has to be discovered and that the discovery cost is the binding constraint; the bull camp’s view is that the discovery is happening on the second axis already and that the constraint is funding it. Both views are consistent with the cost-per-capability data through November 2025. Neither is consistent with the headline that ran the day after the NeurIPS talk. The qualifier was always there.