---
title: "What Sutskever won't say, and how it positions him"
publication: deslop.media
byline: deslop.media editorial
provenance: editorial_synthesis
provenance_note: "Editorial synthesis by deslop.media editorial, assembled from the subject's public record. The subject did not sit for an interview with deslop.media and has not reviewed it."
issue: 1
issue_title: The Scaling Bet
grade: draft
issue_version: 3
issue_version_date: 2026-10-11
piece_version: 1
first_published: 2026-10-11
reporting_current_through: 2026-10
canonical_url: https://deslop.media/feature/ilya-sutskever/the-missing-principle-and-bear-substrate/
cite_as: "\"What Sutskever won't say, and how it positions him,\" deslop.media editorial, Issue 1 (Draft v1, October 11, 2026)"
changes: https://deslop.media/issues/1-the-scaling-bet/changes/
flag: "MCP tool flag_content, or POST https://api.deslop.media/api/flags"
---

Asked in November 2025 what he was looking for, Ilya Sutskever, who co-founded OpenAI and now runs Safe Superintelligence Inc., gave Dwarkesh Patel an answer in the register of a critic, not an engineer. "Ugliness, there's no room for ugliness… beauty, simplicity, elegance, correct inspiration from the brain." Pressed for the specific machine-learning principle SSI was pursuing, he declined: circumstances, he said, made it hard to discuss in detail.

That is the most specific public statement Sutskever has made about the technical bet his $32 billion lab is funded to pursue. He has not published a paper. He has not shipped a model. SSI's website carries a manifesto and three update posts, the newest a July 26, 2026 announcement of its NVIDIA partnership, and none of them a technical publication. Beyond the aesthetic criteria and a couple of analogies, the principle SSI is built on is, by his own description, a withheld claim.

Sutskever has named what he rejects — the "pretraining as we know it" recipe, the assumption that 100× scale closes the gap — and several researchers who reject roughly the same thing now have funded labs of their own. Sutskever sits inside that group. He sits in a particular position within it. The position has shape, even if the principle does not.

## What he has hinted at

Read carefully, the Dwarkesh November 2025 interview gives three topical signals that are not the principle itself but constrain what it might be.

The first is generalization. "These models somehow just generalize dramatically worse than people. It's super obvious. That seems like a very fundamental thing." His vibe-coding example — the model fixes a bug, introduces a second bug, then reintroduces the first when told about the second — is offered as the gap between benchmark scores and real-world economic impact, what he calls "jaggedness." The diagnosis is that brute-force optimization on evaluations does not produce taste or transfer.

The second is value functions. Sutskever invokes a patient described by the neuroscientist Antonio Damasio, a man who after a brain injury could deliberate but not feel, to suggest that human values are evolutionarily hard-coded shortcuts that current ML lacks. The two-students analogy makes the point in pedagogical form: a 10,000-hour competitive programmer who has memorized everything, set against a 100-hour student who has "it." The first wins benchmarks; the second is what generalization means. The contrast is at the level of training process, not architecture.

The third is continual learning. Sutskever described a possible SSI system as "a superintelligent 15-year-old that's very eager to go," sent out to learn a job, and said "the deployment itself will involve some kind of a learning trial-and-error period." That is the closest the interview comes to a mechanism. It is not specified.

These signals do not name a substrate. Sutskever does not endorse or reject world models. He does not endorse or reject state-space architectures. He does not endorse or reject pure reinforcement-learning approaches. He does not say anything about transformers versus alternatives. The hints are at the level of objective and generalization; the substrate is left blank.

## The camp around him

Since 2024, a research camp has accumulated a recognizable shape and a recognizable cap table around the recipe-rejection. Every other named participant has put a substrate on the table. Sutskever has not.

Yann LeCun, who spent twelve years at Meta — including roles as FAIR founding director and chief AI scientist — and left at the end of 2025, has named the substrate in print. "Joint Embedding Predictive Architecture" — "JEPA" in his shorthand — is his proposed alternative to autoregressive language modeling: a non-generative system that learns to predict latent representations of the world from video. V-JEPA 2, published as arXiv:2506.09985 in June 2025, was pre-trained on more than a million hours of internet video; the action-conditioned variant V-JEPA 2-AC was deployed zero-shot on Franka robot arms in two different labs. At AI House Davos on January 23, 2026, LeCun's standing position got its sharpest formulation: "LLMs basically are a dead end when it comes to superintelligence." On March 10, 2026, his AMI Labs raised $1.03 billion at a $3.5 billion pre-money valuation, the largest seed round in European startup history, co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions, with NVIDIA, Samsung, and Temasek participating.

Richard Sutton, the reinforcement-learning pioneer at the University of Alberta, has named a different one. Sutton's "OaK" architecture — feature construction, subtask, option, model, planning — was presented at RLC 2025 and NeurIPS 2025; it is model-based RL with continual learning, online cross-validated step sizes, and option discovery. In his September 26, 2025 Dwarkesh interview, Sutton said LLMs are a dead end on the question of learning from experience, and he rejected the bull reading of his own "Bitter Lesson." That essay, often invoked as scripture for the scale-pilled, was about general methods that scale with compute beating human-knowledge-engineered approaches. Sutton clarified that LLM pretraining, in its current form, does not qualify as the kind of method he had in mind.

Albert Gu, the Carnegie Mellon researcher behind Mamba, has named a third. State-space models. His Mamba-3 paper, posted as arXiv:2603.15569 on March 16, 2026, argues that on a flop-for-flop basis, state-space architectures match transformers at language modeling, with deficits on tasks requiring exact recall. Production hybrids are now shipping at scale: NVIDIA's Nemotron-H in April 2025, Nemotron Nano 2 in August 2025, the Nemotron 3 family in December 2025, IBM Granite 4.0 in October 2025, AI21's Jamba 1.5. Not at frontier-flagship scale yet, but at the open-weights frontier.

Sutskever's bet sits at a higher valuation than the named-substrate bets — $32 billion versus AMI Labs' $3.5 billion pre-money — and with far less of a public technical record. He has the recipe-rejection. He does not have the substrate.

<!-- figure:primary -->

## The pattern

What was a Sutskever idiosyncrasy in 2024 — the man with the famous departure, the unusual fundraise, the no-product mandate — is now a pattern. The three other bettors with named architectures — LeCun, Sutton, Gu — are pursuing alternatives to the LLM-pretraining-recipe-as-it-stands, with funding that ranges from university labs and startups (Sutton's Oak Lab, Gu's Cartesia) to billion-plus venture (LeCun's AMI Labs). The 2025 AAAI Presidential Panel on the Future of AI Research, surveying 475 AI researchers, about two-thirds of them academics, found that 76 percent assert scaling current approaches to AGI is "unlikely" or "very unlikely" to succeed. Academic majority and frontier-lab funding now point at the same recipe-rejection.

Inside that group, Sutskever's position has a particular signature: he has rejected the recipe about as sharply as any other named bettor while saying less about what the substrate alternative is. Every other named bettor has surfaced a paper, a release, a deployed system, a named architecture. Sutskever has surfaced an aesthetic criterion. The other camp members have explicit views on whether the next paradigm will be transformer-shaped, world-model-shaped, state-space-shaped, or RL-pure-shaped. Sutskever's public statements do not foreclose any of those options.

## Why he might not say

One reading is competitive discipline. SSI's secrecy posture is unusual even for a frontier lab — Faraday-cage phone storage for candidates, no interviews from non-Sutskever staff, no model previews, no demos. Sutskever's July 3, 2025 X post, in which he rebuffed Meta's acquisition rumors with the line "we have the compute, we have the team, and we know what to do," is consistent with a lab that has converged on a direction it does not want competitors to map. On this reading, the silence is not vagueness but the deliberate withholding of a specific principle. Publishing it would surrender first-mover advantage on the only thing SSI is funded to produce.

A second reading is a portfolio hedge. Roughly fifty employees and the $3 billion SSI raised before NVIDIA's reported July 2026 investment do not, on their face, force a single architectural commitment. SSI may be running parallel research bets and refusing to disclose any of them precisely so that none of them gets prematurely declared. Public material does not show whether SSI's bet is single-architecture or research-portfolio.

A third reading, the one most consistent with the Sutskever record, is that the principle exists at a layer above architecture. The public framing has been at the level of objective and generalization, not substrate: the bug-loop diagnosis points to what the model is trying to do, not what circuits it runs on; the two-students analogy distinguishes a 100-hour student "with it" from a 10,000-hour grinder at the level of training process; the Damasio reference treats human values as evolutionarily hard-coded shortcuts at the level of objective function. None of those is an architecture choice.

The "as we know it" qualifier from the December 2024 NeurIPS talk runs in the same direction. Sutskever did not say compute would stop working. He said the specific recipe of web-scale autoregressive pretraining on natural text could not be extended further on data alone. Taken with the rest of the interview, that points toward scaling against new objectives: different supervision signals, better generalization, something like the value functions he discussed with Patel. Each of those concerns what the model is trained to do. None is a substrate move.

On this reading, Sutskever has not picked a substrate because the bet does not turn on substrate. LeCun has bet that JEPA is the right circuit; Gu has bet that state-space sequence operators are. Sutskever's bet is that whatever the substrate the empirical work justifies, the model has to be trained against an objective formulation that produces the kind of generalization a human "with it" exhibits — taste, transfer, the ability to hold two failure modes in tension without losing either. The architectures are the implementation of that objective, not the bet itself. On this reading, SSI's silence on architecture matches where the wager sits: at a layer that does not yet have a public artifact because what would be published is not "JEPA versus state-space" but a different way of formulating what the model is trying to learn — and that formulation is, by Sutskever's own description, a withheld claim. The bet survives any of multiple substrate outcomes if the objective formulation is right; it fails on all of them if it is not.

A fourth reading, less charitable, is that there is no principle yet — that what Sutskever has identified is a research direction rather than a result, and that publishing the direction without the empirical validation would invite premature critique. That SSI's own site listed no paper or technical report as of October 10, 2026 is consistent with this reading too.

Each reading is consistent with the public record; none is settled by it.

## What disagrees

The strongest counter to the recipe-rejection view does not come from Sutskever; it comes from the labs that are still shipping.

Demis Hassabis, then Google DeepMind's chief executive and since August 2026 its chair and Alphabet's chief scientist, said at the Axios AI+ Summit in December 2025 that "the scaling of the current systems, we must push that to the maximum… It could be the entirety of the AGI system"; at Davos he placed AGI five to ten years out, a horizon his July 14, 2026 essay shortened to "probably only a few short years away"; and he has continued to ship Deep Think models: an advanced version of Gemini Deep Think reached gold-medal standard at the 2025 International Mathematical Olympiad with 35 of 42 points in July 2025, and Gemini 3 Deep Think reached 84.6 percent on ARC-AGI-2's Semi-Private set, as verified by ARC Prize, in February 2026. Dario Amodei, the Anthropic CEO, told Dwarkesh on February 13, 2026 that his "top level model of scaling is the same as it was in 2017" and that he was at 90 percent confidence on a "country of geniuses in a datacenter" within a decade. Sam Altman's "Three Observations" defends scaling laws as "accurate over many orders of magnitude." Noam Brown, OpenAI's vice president of research, has restated the thesis around major OpenAI releases, most recently in April 2026. The April 7, 2026 Anthropic-Google-Broadcom deal for approximately 3.5 gigawatts of TPU capacity online from 2027, plus about $720–745 billion of 2026 capital spending guidance at Microsoft, Alphabet, Amazon and Meta (total capex, each by its own definition), is the market price of the bull view that the current substrate, with reasoning attached, gets there.

The recipe-rejection response is that the benchmark trajectory does not settle the question. ARC-AGI-3 opened in March 2026 with frontier models below 1 percent against a human baseline of 100 percent. In ARC Prize's July–August 2025 preview agent competition, scored on three private games, the top purpose-built agent, "StochasticGoose" from Tufa Labs, had reached 12.58 percent. On September 2, 2026, though, ARC Prize reported GPT-6 Astra at 62.7 percent on its Semi-Private set under the Standard harness at max reasoning effort. Sara Hooker, co-founder and chief executive of the startup Adaption and formerly head of Cohere's research lab, argues in "On the Slow Death of Scaling" (SSRN, December 2025) that the relationship between training compute and performance is "highly uncertain and rapidly changing." The recipe-rejection reading of GPT-5's August 2025 launch — a router-of-models architecture rather than a single scaled monolith — is that monolithic scaling has plateaued and the labs are quietly hybridizing.

Where Sutskever specifically stands in this argument is partly readable from the December 2024 NeurIPS talk. His framing of AlphaZero — that reasoning makes models more unpredictable, not less — aligns more cleanly with reinforcement-learning histories than with LeCun's world-model bet. He has made no public statement about LeCun's program through May 2026, despite their meta-position alignment.

## What would settle it

The cleanest test of where Sutskever sits is also the simplest one: what does SSI publish, and when? Through May 2026, the answer was nothing: no paper, no model, no blog post beyond the founding manifesto and two update messages. The third update, on July 26, 2026, announced the NVIDIA partnership; it was not a technical publication. The lab's public technical record is a Reuters quote about a "new mountain to climb" and an aesthetic criterion in a podcast. If the principle is competitive discipline, the publication will arrive when SSI has converged enough to defend the priority claim. If the principle is a portfolio, the public artifact may never resolve to a single architectural bet. If the principle is at the layer of objective rather than substrate, the first SSI paper will read more like a generalization-and-value-function paper than a transformer-replacement paper. If there is no principle yet, the first SSI paper will arrive late or not at all.

The camp around him is now legible. LeCun has named "JEPA," shipped V-JEPA 2, and raised $1.03 billion to scale it. Sutton has named "OaK" and presented it at two conferences. Gu has named state-space models and watched them ship in production at NVIDIA, IBM, and AI21. Each has put a named architecture on the public record, where it can be checked.

Sutskever has done none of that, and the position he occupies — at the camp's high-rejection, low-specificity corner — is the one position the camp's pattern does not yet make legible. What he won't say is the only public signal of what he is looking for. The criterion is "beauty, simplicity, elegance, correct inspiration from the brain." The principle is undisclosed. The lab is funded to produce something that is either the principle or its absence. SSI's own site listed no paper, model or technical report as of October 10, 2026. That is the test.