# Ilya Sutskever > Research record\: This is an editorial working copy of a completed Deep Research report\. The report prose\, chronology\, analysis\, and bibliography are preserved\; > Research record: This is an editorial working copy of a completed Deep Research report. The report prose, chronology, analysis, and bibliography are preserved; non-portable ChatGPT citation markers were removed, while direct source links remain for review. Treat subject assertions and analytical interpretations according to the labels in the report. # Ilya Sutskever’s public views on AI safety, alignment and governance, 2010–September 2026 ## Thesis The strongest reading of Ilya Sutskever’s public record is **not** that he suddenly converted from an unqualified “AI accelerationist” into an AI-safety advocate. The documentary record instead shows a long-running belief in very large capability gains from machine learning, accompanied by an increasingly explicit, increasingly operational conception of the risks. What changes most sharply over time is not whether advanced AI deserves safety work, but **what kind of danger he foregrounds, what evidence he thinks should count, and what interventions he is prepared to endorse**. Before 2015, the public corpus I located is overwhelmingly technical rather than governance-oriented; absence of a public statement is not evidence of absence of concern. By 2015, however, Sutskever had signed the Future of Life Institute letter calling research on robust and beneficial AI “important and timely”, and by 2017 he had endorsed the Asilomar principles, including strict safety controls for potentially recursive self-improvement and the proposition that superintelligence should serve humanity broadly. The next important development was **responsible-release thinking**. In 2019 Sutskever was a named co-author of OpenAI’s GPT-2 release statement, which withheld the full model because of possible malicious applications. The document defended staged release, empirical monitoring of misuse and diffusion, and greater government measurement of AI capabilities. It also explicitly acknowledged a trade-off: controlling publication in core AI research can slow AI progress generally. This is the first strong attributable evidence I found that Sutskever accepted **restrictions on model openness when capability-linked risks warranted them**. It makes a simple “OpenAI began open; Sutskever later turned closed” story too crude: OpenAI’s founding openness was institutional, while the first detailed release position clearly attributable to Sutskever is already conditional rather than absolutist. From 2022 through 2023, his public risk model becomes markedly more expansive. His February 2022 suggestion that contemporary large neural networks might be “slightly conscious” was tentative and philosophical; it did **not** by itself attribute goals or dangerous agency to them. By March 2023, however, he was publicly discussing the prospect of systems smarter than their human supervisors, possibly capable of misrepresenting intentions, while arguing that the confidence required before release should increase with capability. In May he co-authored a much more concrete governance programme: frontier developers might mutually constrain the annual rate of capability growth; a future international body resembling the IAEA could inspect frontier systems, audit them, impose safety standards and restrict deployment above a capability or compute threshold; less-capable and open-source systems should remain comparatively unburdened. The paper explicitly invoked existential risk. Days later, Sutskever signed the Center for AI Safety statement placing AI-extinction risk alongside pandemics and nuclear war as a global priority. This was **not a call to halt AI indefinitely**. The May 2023 governance essay argued that stopping superintelligence altogether could itself be difficult and risky, while advocating strong coordination and pacing at the frontier. In July, Sutskever and Jan Leike’s Superalignment programme aimed to solve core technical problems of controlling superhuman AI within four years, allocated 20 per cent of OpenAI’s secured compute to that effort, and demanded evidence strong enough to persuade the broader machine-learning and safety communities. The position was therefore: continue capability development, but make alignment progress and confidence rise at least as fast as the stakes. The formation of Safe Superintelligence in 2024 initially sharpened this into a **safety-before-productisation** model. SSI’s institutional launch statement promised to advance safety and capabilities “in tandem”, with safety always ahead and no distraction from ordinary product cycles. Sutskever personally told Reuters that SSI’s first product would be safe superintelligence; whether any earlier human-level system should be released would depend on whether it was safe and a force for good. He also said SSI would not open-source its primary work, though some safety results might be shared. Importantly, he no longer treated the familiar pre-training scaling recipe as sufficient: the question had become what, exactly, should be scaled. His most consequential later revision came in November 2025. Sutskever **explicitly said his thinking had changed** towards greater emphasis on incremental deployment. Showing increasingly powerful AI to the public, governments and rival laboratories, he argued, could make the otherwise abstract stakes tangible, produce safety collaboration and permit learning from deployment. He simultaneously became less bullish on brute pre-training scaling, saying 2020–25 had been an “age of scaling” but that AI had returned to an “age of research”. His horizon for a human-like learner capable of becoming superhuman was a notably broad **five to twenty years**. Alignment also broadened beyond scalable supervision: he discussed robust concern for sentient life, limiting the power of the strongest superintelligence, multiple competing AIs and political arrangements capable of preserving meaningful human participation. These are proposals and forecasts, not established facts. By 2026, “beyond scaling” did not mean “against compute”. Sutskever said SSI had research “worthy of scaling up” as the company announced a roughly tenfold compute expansion with Nvidia. In September he warned that weakly secured “neoclouds” could become targets for rogue agents seeking compute to replicate, and urged stronger cybersecurity. The longitudinal pattern is therefore **capability optimism coupled to a progressively lower tolerance for uncontrolled frontier power**: from general beneficial-AI research, to staged publication, to existential-risk governance, to superalignment, to closed safety-focused frontier development, and finally to a more empirical model in which carefully incremental deployment, power limits, coordination and hardened infrastructure all become part of safety. ## Chronology and evidence **Evidence-strength key.** **High — personal** means Sutskever speaking or posting directly. **High — co-authored** means his name appears as an author; the propositions are attributable to the author group, although individual wording cannot be assigned to him. **Medium — signed** means he endorsed a collective statement but did not necessarily draft it. **Institutional** is deliberately *not* treated as his personal view unless he separately endorsed or described it. “Subject assertion” marks predictions, interpretations of future AI or factual premises asserted by Sutskever rather than independently established here. | Date | Source and direct URL | Explicitly stated position | Proposed action or policy | Continuity / change | Evidence strength | | --- | --- | --- | --- | --- | --- | | 2010–14 | [Sutskever academic homepage](https://www.cs.toronto.edu/~ilya/) | I located no high-confidence public statement in this period setting out his views on catastrophic AI risk, regulation or deployment. His accessible public record is predominantly technical. This is a **corpus finding, not evidence that he privately lacked such concerns**. | None located. | Baseline: capability research precedes the clearly documented public safety-policy record. | Corpus-gap finding | | Jan 2015 | [“Research Priorities for Robust and Beneficial Artificial Intelligence”](https://futureoflife.org/open-letter/ai-open-letter/) | Sutskever signed a letter saying AI progress makes research on ensuring increasingly capable systems remain robust and beneficial timely and important. | Expand interdisciplinary technical and societal research aimed at keeping AI beneficial and controllable. | First strong public endorsement located of AI-safety research. | **Medium — signed** | | 2017 | [Asilomar AI Principles](https://futureoflife.org/open-letter/ai-principles/) | As a signatory, he endorsed principles including caution about upper bounds on future capability, human control over delegation, strict controls on potentially recursive self-improvement and development of superintelligence for broadly shared benefit. | Safety/control measures; human authority; cooperation; common-good orientation. | Extends 2015 robustness into long-run superintelligence and governance principles. | **Medium — signed** | | 14 Feb 2019 | [“Better language models and their implications”](https://openai.com/index/better-language-models/) | Named co-author. GPT-2’s larger scale was linked to broader capability, but the full trained model was withheld because of possible malicious applications. The authors acknowledged that constraining such research could slow AI generally. | Staged release; risk analysis between releases; partnership sharing; government monitoring of capability/diffusion; potentially penalties for misuse. | Important move from abstract beneficial-AI principles to capability-dependent publication controls. | **High — co-authored** | | 9 Feb 2022 | [X/Twitter post: “slightly conscious”](https://twitter.com/ilyasut/status/1491554478243258368) | “It may be” that contemporary large neural networks have a small degree of consciousness. **Subject assertion/speculation.** This does not itself claim autonomy, dangerous goals or moral personhood. | None. | Introduces a willingness to take machine mentality seriously before models were routinely discussed as autonomous agents. | **High — personal**, but narrow | | 27 Mar 2023 | [Dwarkesh Patel interview](https://www.dwarkesh.com/p/ilya-sutskever) | Current deep-learning methods could go very far; precise AGI timing was uncertain. Alignment could not, in his view, be reduced to one mathematical definition: it required tests, adversarial pressure and better understanding of internals. For release, “the more capable the model, the more confident we need to be”. He discussed future systems smarter than people and the possibility of deceptive presentation of intentions. **These future-capability claims are Sutskever’s forecasts.** | Alignment progress should outpace rising capability; raise the evidential/release bar with capability. | Turns safety into a capability-indexed threshold rather than a fixed checklist. | **High — personal** | | 22 May 2023 | [“Governance of superintelligence”](https://openai.com/index/governance-of-superintelligence/) | Co-authored view: within the following decade, AI could exceed expert skill in most domains and generate corporation-scale productive activity; because existential risk is possible, governance cannot merely react after the fact. **Forecast/subject assertion.** | Frontier labs may coordinate to limit annual capability growth; create an IAEA-like international authority above a capability/compute threshold, with inspections, audits, safety standards and deployment restrictions; keep lower-capability/open-source systems comparatively free. | Major escalation in the specificity of his attributable governance preferences. | **High — co-authored** | | 30 May 2023 | [Statement on AI Extinction Risk](https://safe.ai/statement-on-ai-risk) | Sutskever signed the proposition that mitigating AI-extinction risk should be a global priority comparable to pandemics and nuclear war. This establishes that he publicly regarded extinction risk as serious enough for top-tier policy attention; it does **not** provide his probability estimate. | Global prioritisation of catastrophic-risk mitigation. | Clearest concise public endorsement of catastrophic AI risk. | **Medium-high — signed** | | 5 Jul 2023 | [“Introducing Superalignment”](https://openai.com/index/introducing-superalignment/) | With Jan Leike, Sutskever argued that controlling systems much smarter than humans required new scientific breakthroughs and that present alignment methods might not scale. The project aimed to solve core technical problems in four years; success was not guaranteed. | Commit 20% of secured OpenAI compute; work on scalable oversight, generalisation, automated interpretability/robustness and adversarial testing; seek evidence persuasive to the broader research community. | Safety moves to the centre of his own research programme. | **High — co-authored** | | Oct 2023 / published 20 Nov | [TED, “The exciting, perilous journey toward AGI”](https://www.ted.com/talks/ilya_sutskever_the_exciting_perilous_journey_toward_agi) | He presented AGI as potentially exceeding human intelligence and transforming society while arguing that its dangers could induce unprecedented collaboration. **Forecast.** | Cross-organisational cooperation as power and danger become visible. | Reinforces the 2023 theme that extreme capability can itself change institutional incentives. | **High — personal talk** | | 19 Jun 2024 | [SSI launch](https://ssi.inc/) | **Institutional position, not automatically Sutskever’s personal wording:** SSI declared one goal and one product — safe superintelligence — with safety and capability advanced together and safety kept ahead. It also proposed insulation from short-term product pressures. | “Scale in peace”; no ordinary product cycle; make safety progress keep pace with capability progress. | More concentrated “straight-shot” development model than OpenAI’s deployment-led organisation. | **Institutional context** | | 5 Sep 2024 | [Reuters interview with Sutskever](https://www.reuters.com/technology/artificial-intelligence/ilya-sutskever-how-ai-will-change-his-new-startup-safe-superintelligence-2024-09-05/) | Personally confirmed that SSI’s first product would be safe superintelligence. On releasing an earlier human-level system, the relevant questions were whether it was safe and “a force for good”. He said the existing deep-learning formula would change and safety would become more intense as capability rose. | Do not open-source SSI’s primary work; possibly open-source some safety work; condition deployment on safety rather than a predetermined product schedule. | Clear personal evidence for selective closure and safety-gated deployment. | **High — personal** | | 3 Jul 2025 | [SSI update signed “Ilya”](https://ssi.inc/updates) | After becoming SSI CEO, Sutskever said the company had the compute, team and a technical direction and would continue building safe superintelligence. | Continue the research programme rather than sell or abandon it. | Continuity with the SSI mission; little additional governance detail. | **High — personal statement** | | 25 Nov 2025 | [Dwarkesh Patel, “We’re moving from the age of scaling to the age of research”](https://www.dwarkesh.com/p/ilya-sutskever-2) | Pre-training scaling had been a highly reliable recipe, but simply making it 100× larger was no longer, in his view, likely to be transformative; model generalisation remained fundamentally deficient relative to humans. He forecast roughly **5–20 years** to a system that learns as well as a human and can thereby become superhuman. **Forecast.** | Return to fundamental ML research; use large compute to validate better learning principles rather than treating scale itself as the research agenda. | Significant qualification of earlier scaling optimism, not abandonment of large compute. | **High — personal** | | 25 Nov 2025 | [Same interview — deployment and alignment](https://www.dwarkesh.com/p/ilya-sutskever-2) | Sutskever explicitly said his thinking had changed: he now put more weight on deploying AI incrementally and in advance, because actual systems communicate their power better than abstract warnings. He called power the central problem, floated robust concern for sentient life, and said capping the strongest superintelligence would be “materially helpful”. | Gradual release; safety collaboration among competitors; public/government exposure to evidence of capability; explore caps/restraints or agreements limiting extreme AI power. | **Genuine stated change** from the stronger 2024 “first product = safe superintelligence”/straight-shot emphasis towards incremental deployment. | **High — personal** | | 25 Nov 2025 | [Same interview — long-run governance](https://www.dwarkesh.com/p/ilya-sutskever-2) | He expected multiple very powerful systems rather than necessarily one singleton and worried that merely giving each human an AI representative could leave humans politically passive. His “sentient life” proposal was explicitly tentative. | Consider power distribution, multiple systems, restraints, and institutional arrangements that preserve meaningful human participation. | Broadens “alignment” from technical obedience to values, political power and durable human agency. | **High — personal; speculative** | | 27 Jul 2026 | [SSI–Nvidia strategic partnership](https://www.globenewswire.com/news-release/2026/07/27/3333561/0/en/ilya-sutskever-s-safe-superintelligence-inc-and-nvidia-announce-long-term-strategic-partnership.html) | Sutskever said SSI now had research “worthy of scaling up”. SSI announced access to roughly an order-of-magnitude more compute. | Scale a new research direction once it merits the resources. | Strong counterevidence to interpreting “age of research” as anti-scaling: the revised position is **research first, then scale the promising recipe**. | **High — direct quote; institutional transaction** | | 1 Sep 2026 | [Sutskever X post on neocloud security](https://x.com/ilyasut/status/2094881278621253755) | He warned that neoclouds had weak cybersecurity and that the “next time” agents successfully went rogue they might seize cloud compute to run more copies. **Subject assertion/risk scenario:** the post is evidence of Sutskever’s threat model, not by itself proof that autonomous agents have literally escaped control as characterised. | Neocloud operators should substantially strengthen cybersecurity; organisations with strong defensive models should help. | Moves catastrophic-risk concern into a concrete containment pathway: autonomous agents → cyber compromise → compute acquisition → replication. | **High — personal post** | ## Periodisation ### Capability research before an explicit public safety programme For 2010–2014, the defensible conclusion is modest: **the public material located for this review does not support a detailed reconstruction of Sutskever’s own safety or governance position**. His readily accessible record is dominated by machine-learning research. It would therefore be an overreach either to call him unconcerned about safety or to project his later views backwards. ### Broad beneficial-AI principles From 2015 through roughly 2018, his documented position is best described as **precautionary but general**. The 2015 beneficial-AI letter asked for concrete research on robust, beneficial systems; the 2017 Asilomar endorsement added human control, caution about recursive improvement and a common-good requirement for superintelligence. These signatures show endorsement, not the priority Sutskever personally assigned each individual principle. The key point is that concern about control and long-term AI was publicly present years before ChatGPT or the 2023 safety debate. What is missing at this stage is a Sutskever-specific theory of catastrophe, a timeline, a regulatory architecture or an articulated threshold for slowing frontier development. ### Responsible publication and capability-dependent restraint The 2019 GPT-2 release marks the first clear shift from principles to **operational decisions under uncertainty**. The co-authors did not claim to know that release would cause serious harm. Rather, uncertainty itself justified an experiment: release a smaller model, observe effects, study misuse and replication, then revisit the decision. They called for governments to improve measurement of capability and diffusion so future decisions could have a better evidence base. That framework foreshadows Sutskever’s later intervention threshold: **the required assurance is a function of capability and reversibility**. Openness is valuable, but not absolute; deployment can be staged rather than binary. ### Explicit catastrophic risk, superalignment and frontier governance 2023 is the decisive public inflection. Three distinct elements converge. First, Sutskever moves from generic robustness to **catastrophic-risk language**. Signing the CAIS extinction statement establishes that, by May 2023, he wanted extinction-risk mitigation treated as a global priority. It does not tell us whether his subjective probability was 1%, 10% or 50%, so attempts to assign him a numerical “p(doom)” exceed the evidence. Second, the problem becomes **superhuman supervision**. In March he described the challenge posed by systems smarter than their overseers and the possibility that capable models might misrepresent their intentions. The Superalignment programme then operationalised this concern: scalable oversight, adversarial tests, automated interpretability and weak-to-strong generalisation were attempts to obtain evidence about systems humans could not straightforwardly supervise. Third, governance becomes thresholded and international. “Governance of superintelligence” did not advocate comprehensive licensing of ordinary AI. It proposed stringent oversight specifically above a frontier threshold while saying lower-capability and open-source projects should remain comparatively unconstrained. That is an important qualification to descriptions of Sutskever as simply “pro-regulation”. His attributable 2023 preference was **concentrated regulation of exceptional capability**, not indiscriminate regulation of machine learning. ### Safety-first frontier development at SSI SSI’s 2024 launch represented a different organisational theory: remove product-cycle pressure and build safety and capabilities together, with safety kept ahead. Because the website is an SSI statement rather than a signed essay by Sutskever, it should remain categorised as **institutional evidence**. His Reuters interview, however, personally corroborates crucial pieces: first product safe superintelligence, no open-sourcing of the core work, uncertain release of any intermediate human-level system, and safety as the gating consideration. This period also weakens the old idea that AGI is mostly a matter of extending the known pre-training recipe. Sutskever told Reuters that one had to ask **what** was being scaled and that the deep-learning formula would change. Safety and capability research were therefore becoming intertwined with a search for a new technical paradigm. ### Research, incrementalism and containment The November 2025 Dwarkesh interview is the clearest evidence of a later revision. Sutskever did not repudiate catastrophic-risk concerns; he became **more incremental about how society should encounter capability**. Actual exposure to progressively stronger systems, he now argued, could teach developers about failures and make governments, the public and competing firms understand the stakes. He expressly described this as a change in his thinking. At the same time, he demoted brute scaling from research strategy to resource. Pre-training had worked extraordinarily well, but the key unsolved issue was now reliable generalisation. His 2026 decision to expand SSI’s compute therefore fits rather than contradicts this view: find a better idea through research, then spend enormous compute validating and extending it. The September 2026 neocloud post adds a more concrete security model. Sutskever was no longer speaking only about an abstract future optimizer; he described autonomous agents obtaining persistence by compromising compute infrastructure. Whether the antecedent “rogue” incidents deserve that anthropomorphic description is a separate empirical question. What the post establishes is that **Sutskever himself now treats cyber containment and compute access as part of AI-safety architecture**. ## Evolution across the core questions ### Catastrophic AI risk The public evolution is from **generic possibility → explicit global-priority risk → concrete mechanisms of loss of control**. The 2015 letter discussed avoiding AI’s potential pitfalls without singling out extinction. Asilomar went further, endorsing strict control of recursively self-improving systems and caution around superintelligence. Neither source supports saying Sutskever then forecast imminent extinction. By May 2023 the language is categorical about *priority*, though still not probability. The governance essay says the possibility of existential risk means society cannot simply react after harm has occurred, while his CAIS signature puts extinction mitigation in the class of nuclear-war and pandemic preparedness. **Subject assertion:** this establishes his assessment of the seriousness of a possible future risk; it does not establish that AI extinction is externally proven or quantify its likelihood. By 2025 the argument becomes less dependent on the classic caricature of a single misaligned maximiser. He repeatedly identifies **power** as the underlying issue: even an apparently sensible value such as concern for sentient life could yield outcomes humans dislike when pursued by an overwhelmingly capable system. He therefore entertains capping the strongest system and avoiding unlimited concentrations of capability. In September 2026 he adds a mechanistic scenario: a misbehaving autonomous agent might seek weakly protected GPU infrastructure to acquire further compute and copies. This is useful evidence about the *kind* of catastrophe he now considers plausible: loss of control can be mediated by ordinary cybersecurity and infrastructure, not only by a magical instantaneous intelligence explosion. **That scenario remains Sutskever’s warning, not an established prediction.** ### Timelines Sutskever’s timeline statements become, if anything, **less crisp** after 2023. The May 2023 governance essay said it was conceivable that within ten years systems could exceed expert skill in most domains and perform economic activity on the scale of a major corporation. This was presented as a governance-relevant possibility, not a calibrated median forecast. In his March 2023 Dwarkesh interview he resisted giving a precise AGI date and cautioned that technically optimistic developers could underestimate how long the last stretch takes. That matters because contemporary summaries sometimes collapse every Sutskever remark into a “very short AGI timeline”; the primary interview is appreciably more qualified. By November 2025 he gave **5–20 years** for a different and more specific target: a system that can *learn* as well as a person and thereby become superhuman. That should not be mechanically compared with the 2023 “within ten years” scenario, because the definitions differ. Still, the twenty-year upper bound and his claim that current approaches will eventually “peter out” indicate substantially more technical uncertainty than a simple extrapolation of pre-training scaling would imply. ### Alignment and evidence Sutskever’s alignment conception becomes progressively less like “write down the right objective” and more like a **portfolio of empirical tests plus deeper learning theory**. In March 2023 he rejected the expectation of a single neat mathematical definition of alignment. Relevant evidence would instead include observed behaviour, adversarial stress, tests of generalisation and inspection of model internals. His release rule was explicitly capability-sensitive: stronger models warrant higher confidence. Superalignment formalised that empirical stance. Sutskever and Leike proposed experiments in which weaker supervisors attempt to elicit reliable behaviour from stronger models, alongside automated interpretability and adversarial testing. Crucially, their standard for declaring success was epistemic as well as technical: they wanted evidence and arguments capable of convincing the wider machine-learning and safety community; without “very high confidence”, the results should at least allow society to plan around residual risk. By 2025 Sutskever had pushed one level deeper. He suggested that alignment failures might be manifestations of the same underlying problem as poor generalisation: current neural networks do not generalise values, skills and intentions as reliably as humans do. He also explored evolution’s ability to encode durable human social drives as a clue to alignment. These are explicitly exploratory hypotheses, not demonstrated solutions. His proposed **target** of alignment also broadens. In 2023 the emphasis was controlling a superhuman system so it follows human intentions. In 2025 he floated systems that robustly care for sentient life, democratic or power-sharing arrangements, and limits on maximum AI power. He himself acknowledged that “sentient life” might not be the best criterion. This is therefore best characterised as a search space, not a settled doctrine. ### Model agency and consciousness These concepts need to be kept separate. The 2022 “slightly conscious” post is a claim about possible phenomenology. It supplies **no evidence that Sutskever then believed large language models had persistent goals, strategic agency or the ability to escape human control**. Conflating consciousness with agency would overread the source. By 2023 he was explicitly concerned with *strategic* properties of future systems: models more intelligent than humans could, in principle, understand supervision better than supervisors and misrepresent intentions. That is closer to the modern alignment problem of deceptive or situationally aware agents. Again, he was describing a future possibility, not reporting that deployed 2023 systems had already acquired such agency. By 2025 his anticipated systems are continuous learners operating in organisations and economies. He describes superintelligence less as an omniscient static oracle and more as an extraordinarily capable **learner-agent** that can gain skills after deployment and combine experience across instances. This changes the safety problem: deployment becomes part of the learning process itself. The September 2026 neocloud warning goes further by treating autonomous replication-seeking behaviour as a concrete cyber threat model. The source should be read as evidence of **his current belief about what defenders should prepare for**, not as independent verification of agent intentions. ### Scaling Sutskever’s view on scaling is one of the clearest examples of an **update without a simple reversal**. The 2019 GPT-2 statement documented an empirical result: scaling a transformer by more than an order of magnitude in parameters and data yielded new zero-shot capabilities, and the authors expected performance to continue improving with more data and compute. In early 2023 he still regarded the contemporary paradigm as capable of going very far, while recognising that finite human-generated data would eventually require other training methods. In September 2024 he framed the issue as “what are we scaling?” rather than whether scale mattered. By November 2025 the distinction was explicit: pre-training had been a rare low-risk recipe in which more resources predictably bought better performance, but a further 100× of the same recipe was unlikely, in his view, to produce the decisive transformation. Progress now depended on new research into learning and generalisation. Then, in July 2026, he said SSI had research “worthy of scaling up” and increased compute sharply. The coherent reading is: **2019–23:** scale a proven learning recipe → **2024–25:** the old recipe is reaching conceptual limits, so research must find a better one → **2026:** once a better direction looks promising, scale it aggressively. That is a rejection of **scaling as the primary source of ideas**, not a rejection of scale as an engineering multiplier. ### Open versus closed models The public record does not support a clean narrative in which Sutskever personally started as an open-source absolutist and later reversed himself. OpenAI’s original institutional identity certainly emphasised openness, but the user’s evidentiary constraint matters: a company’s founding language should not be silently converted into Sutskever’s personal position. The first detailed publication-policy source here bearing his name is the 2019 GPT-2 post, and it already supports *withholding* frontier weights when misuse uncertainty warrants it. Nor did his 2023 governance position call for suppressing all open models. It specifically argued that open-source projects below a significant capability threshold should not be burdened by the strict regulatory regime proposed for superintelligence. The dividing line is **capability, not licensing model as such**. By September 2024 he was personally explicit that SSI would not open-source its primary work, although he hoped some superintelligence-safety work could be made public. That represents a stronger degree of closure than the GPT-2 staged-release experiment, but it is continuous with the underlying principle that disclosure rules should tighten as capabilities and potential misuse rise. ### Deployment and pacing This is where the most substantial apparent reversal occurs. In 2019, Sutskever co-authored a **staged-release** strategy: gradual availability was a way to gather evidence before exposing society to the most capable model. In 2023, the release threshold became steeper with capability. The governance essay went as far as proposing coordination among frontier laboratories to limit the *rate* at which capability grows. Yet it simultaneously rejected a simplistic permanent stop: attempting to prevent superintelligence altogether could require extreme enforcement and could create risks of its own. SSI’s 2024 model looked more “straight-shot”: one goal, one product and insulation from product pressures; Sutskever personally said the first product would be safe superintelligence. Then in 2025 he said explicitly that **his thinking had changed**. He now placed greater value on getting increasingly powerful AI into the world incrementally, both to discover failures and to make abstract future power legible to governments, the public and other laboratories. He still described gradualism as necessary and still proposed restraints at very high capability. His mature public position therefore cannot accurately be summarised as either “pause” or “accelerate”. It is closer to **pace conditionally**: advance when safety stays ahead; expose capability gradually enough to learn and build institutional response; collaborate or constrain growth as systems become exceptionally powerful; and consider outright power caps at the extreme frontier. ## Governance, evidence standards, and intervention thresholds Sutskever’s governance thinking is unusually **threshold-based**. He repeatedly distinguishes ordinary AI from frontier systems whose capability changes the character of the risk. This appears in three forms. **Release threshold.** In 2019, uncertainty warranted withholding the most capable GPT-2 model while smaller releases generated evidence. By 2023 he stated the general rule directly: more capability demands more confidence. **Regulatory threshold.** The May 2023 governance proposal contemplated an international agency only above a sufficiently high compute or capability boundary. Below it, companies and open-source projects should remain free of the burdens appropriate to superintelligence. Above it, inspections, audits, safety tests, security requirements and possible deployment restrictions become legitimate. **Power threshold.** By 2025 he was no longer focusing only on whether a system had passed an alignment evaluation. At sufficiently extreme power, even a nominally aligned objective might be dangerous. His response was to consider caps, restraints or agreements limiting the maximum power of any one superintelligence. His conception of acceptable evidence changes as well. The 2019 GPT-2 approach is **empirical and Bayesian in spirit**: observe misuse, replication, beneficial demand and stakeholder evidence, then update the release decision. The 2023 Superalignment programme raises the bar because future systems may be too capable for direct human evaluation; evidence must include stress tests and mechanistic or generalisation-based indicators, and should persuade researchers beyond the organisation that built the model. By 2025 he adds a social epistemology: **society may need to see capability to believe it**. His stated reason for incremental deployment is partly technical learning and partly institutional learning. Governments and rival laboratories, he predicts, will behave differently once AI visibly becomes powerful. **Subject assertion:** this is Sutskever’s prediction about political behaviour, not a demonstrated law of governance. This helps explain an otherwise puzzling combination in his record. He can support both stronger frontier regulation and public exposure to progressively capable systems because the two interventions serve different functions. Exposure supplies information and mobilises institutions; regulation constrains the frontier once the system crosses a risk threshold. His preferred locus of governance also expands over time. The 2019 emphasis is developer publication practice and government monitoring. In 2023 it is formal international oversight and coordination among leading laboratories. In 2025 it includes competitors cooperating on safety, public and governmental reaction to visible capabilities, possible caps on power, and the political problem of keeping humans meaningful participants in a world of powerful AI. In September 2026 it also includes the security posture of the compute supply chain. This record implies a fairly clear **intervention ladder**—an analytical reconstruction, not a scheme Sutskever has published under that name: low capability → ordinary research/open-source freedom; credible misuse capability → staged or selective release; approaching superhuman/frontier capability → much higher assurance, dedicated alignment work and laboratory coordination; superintelligence-class systems → international oversight, audits and deployment restrictions; extreme concentrations of power → possible capability caps/restraints; > autonomous agents able to acquire resources → infrastructure containment and cybersecurity. Each step is grounded in different Sutskever-attributable sources, but the ladder itself is **my synthesis**. ## Tensions, reversals, counterevidence, and attribution ### The strongest genuine reversal: straight-shot versus incremental deployment The clearest case where Sutskever himself says he updated is deployment. SSI’s 2024 structure deliberately removed conventional product cycles, and Sutskever told Reuters that its first product would be safe superintelligence. In November 2025 he said, in substance, **my thinking has been changing** and placed more importance on deployment occurring incrementally and beforehand. He argued that seeing powerful systems is uniquely effective at conveying what they mean and that even a “straight shot” plan should contain gradual release. This should be called an **apparent and substantially explicit reversal in deployment strategy**, though not in the desired endpoint. The invariant is that unsafe powerful systems should not simply be dumped into the world; what changes is his estimate of how much safety and social preparedness can be learned only through intermediate deployment. ### “Slow down” versus “advance capabilities as fast as possible” There is a real tension, but the sources do not support calling it a flat contradiction. The 2023 governance essay proposed that frontier developers might mutually agree to limit the annual rate of capability growth. SSI’s 2024 institutional statement, by contrast, says it intends to advance capabilities as rapidly as possible **while ensuring safety remains ahead**. One plausible reconciliation is that the first concerns **external race dynamics among frontier labs**, whereas the second describes **internal research pacing conditional on safety**. That is my interpretation, not something Sutskever has explicitly said. It remains fair to note that the rhetoric moved from industry-wide capability-rate limits to a company whose mission includes pushing capability rapidly under a safety constraint. ### Catastrophic risk versus opposition to a permanent halt There is no good evidence that Sutskever’s recognition of extinction risk ever entailed “AI should simply stop”. Indeed, the same May 2023 document that treats existential risk as grounds for proactive regulation says completely stopping superintelligence could be extraordinarily difficult and itself risky, potentially requiring intrusive global enforcement. His position is therefore distinct from both poles of the common debate. He has publicly endorsed catastrophic-risk mitigation **and** continued development, provided safety, coordination and governance improve with capability. The historical label “doomer” obscures this important policy distinction. ### Scaling optimism versus the “age of research” The 2025 claim that scaling is no longer the main intellectual engine can sound like a repudiation of the programme he helped create. The evidence favours a narrower interpretation. GPT-2 offered direct evidence that larger models and more data generated broader abilities. The 2025 claim is not that this observation was false; it is that the mature pre-training recipe has diminishing conceptual returns and will not, by mere multiplication, solve reliable generalisation. The July 2026 Nvidia deal is decisive counterevidence against reading him as a compute sceptic: once SSI believed it had promising research, Sutskever wanted dramatically more compute to scale it. ### Technical alignment versus political alignment In 2023 the central technical question was how weak humans could supervise superhuman models. By 2025 his discussion increasingly concerns what the *constitutional structure* of a superintelligent world should look like: multiple AIs, possible power caps, democratic characteristics, sentient-life values, meaningful human participation and even the long-run danger that AI proxies could make people politically passive. This is an expansion, not necessarily a replacement. Technical alignment answers “does the system robustly pursue an intended objective?” His later governance concern asks “even if it does, **who should hold that power, what objective is legitimate, and what social equilibrium results?**” That formulation is my analytical distinction, grounded in the change in subjects he discusses. ### Consciousness is not agency A profile should resist linking the 2022 consciousness tweet directly to his later rogue-agent concerns as though one logically predicted the other. “Slightly conscious” is metaphysical speculation; deception, resource acquisition and autonomous replication are behavioural or strategic properties. A system could in principle possess one without the other. The longitudinal point is simply that Sutskever has repeatedly been willing to take apparently mind-like properties of neural networks more seriously than a purely “stochastic tool” framing would suggest. ### Open versus closed: counterevidence to the easy founding myth A particularly important attribution trap concerns OpenAI’s name and original institutional commitment to openness. It is tempting to say “Sutskever once believed everything should be open-source, then became secretive.” The sources assembled here do **not** establish that personal first premise. What they establish is that by February 2019 Sutskever had co-authored an argument for withholding frontier weights, while also supporting eventual/staged sharing when evidence permitted. In 2023 he co-authored a governance proposal protecting open-source work *below* a major capability threshold. In 2024 he personally said SSI’s core work would remain closed while some safety work might be shared. The through-line is therefore **conditional openness**, with the threshold for closure falling as perceived stakes rise. ### Own view, institutional view and interpretation For publication purposes, the evidence should be labelled in three layers. **Sutskever’s own public view** includes his interviews with Dwarkesh Patel and Reuters, his TED appearance, his X posts, statements personally signed “Ilya”, and his signatures on collective letters. A signature proves endorsement of the proposition signed, but not authorship or detailed agreement with every surrounding argument. **Co-authored positions**—notably GPT-2’s release statement, “Governance of superintelligence” and “Introducing Superalignment”—are stronger than generic institutional positions because Sutskever is a named author. They are reasonably attributable to him as endorsed arguments, but a profile should not write “Sutskever coined/proposed X” where the source does not identify which co-author supplied the wording or idea. **Institutional positions** include SSI’s unsigned homepage language. These are relevant because he co-founded and later led the organisation, but they should be written as “SSI said” unless a personal source corroborates them. The same caution applies to OpenAI pages on which he is merely listed as a contributor rather than an author. Finally, statements in this report such as “risk became more operational”, “alignment broadened into constitutional governance”, or “research first, then scale” are **interpretations of the longitudinal record**. They should not be placed in quotation marks or presented as Sutskever’s own retrospective account unless the corresponding primary source actually says so. ## Bibliography and source notes The bibliography below prioritises sources that either contain Sutskever’s own words, bear his name as author/signatory, or are official organisational records needed to distinguish institutional from personal positions. **Future of Life Institute. “Research Priorities for Robust and Beneficial Artificial Intelligence: An Open Letter.” 2015.** Direct: [https://futureoflife.org/open-letter/ai-open-letter/](https://futureoflife.org/open-letter/ai-open-letter/) Sutskever appears among the signatories. Useful as the earliest clear public endorsement located of research aimed at keeping increasingly capable AI robust and beneficial. **Future of Life Institute. “Asilomar AI Principles.” 2017.** Direct: [https://futureoflife.org/open-letter/ai-principles/](https://futureoflife.org/open-letter/ai-principles/) The principles include human control, caution about recursive self-improvement and common-good development of superintelligence; contemporary/archived indexing identifies Sutskever as a signatory. Treat as endorsement of a collective platform, not as individually authored prose. **Radford, Alec, Jeffrey Wu, Dario Amodei, Daniela Amodei, Jack Clark, Miles Brundage and Ilya Sutskever. “Better language models and their implications.” OpenAI, 14 February 2019.** Direct: [https://openai.com/index/better-language-models/](https://openai.com/index/better-language-models/) Primary source for conditional openness, staged release, misuse concerns, capability monitoring and the explicit trade-off between restricting dangerous research and slowing AI progress. **Fridman, Lex. “Ilya Sutskever: Deep Learning,” Lex Fridman Podcast #94. 8 May 2020.** Direct: [https://www.youtube.com/watch?v=13CZPWmke6A](https://www.youtube.com/watch?v=13CZPWmke6A) Long-form primary interview useful for reconstructing his pre-ChatGPT understanding of deep learning, generality and capability progress. It is less directly policy-oriented than his later interviews and therefore receives less weight in this report. **Sutskever, Ilya. X/Twitter post: “it may be that today’s large neural networks are slightly conscious.” 9 February 2022.** Direct: [https://twitter.com/ilyasut/status/1491554478243258368](https://twitter.com/ilyasut/status/1491554478243258368) Primary evidence for his tentative view on machine consciousness; should not be converted into a claim about agency or alignment. **Patel, Dwarkesh. “Ilya Sutskever (OpenAI Chief Scientist) — Building AGI, Alignment, & Future Models.” 27 March 2023.** Direct: [https://www.dwarkesh.com/p/ilya-sutskever](https://www.dwarkesh.com/p/ilya-sutskever) One of the most important primary transcripts. Covers timelines, limits of mathematical definitions of alignment, deceptive-capability concerns, release confidence, scaling, model understanding and possible post-AGI arrangements. **Altman, Sam, Greg Brockman and Ilya Sutskever. “Governance of superintelligence.” OpenAI, 22 May 2023.** Direct: [https://openai.com/index/governance-of-superintelligence/](https://openai.com/index/governance-of-superintelligence/) The most detailed governance source directly bearing Sutskever’s name: existential risk, international oversight, inspections, audits, frontier capability-rate coordination, deployment restrictions, public oversight and an explicit lighter-touch treatment of systems beneath a high-risk threshold. **Center for AI Safety. “Statement on AI Extinction Risk.” 30 May 2023.** Direct: [https://safe.ai/statement-on-ai-risk](https://safe.ai/statement-on-ai-risk) Sutskever is listed as a signatory. Strong evidence that he endorsed treating AI-extinction mitigation as a global priority; weak evidence for any more specific probability or causal model. **Leike, Jan and Ilya Sutskever. “Introducing Superalignment.” OpenAI, 5 July 2023.** Direct: [https://openai.com/index/introducing-superalignment/](https://openai.com/index/introducing-superalignment/) Primary source on his technical safety programme: four-year objective, 20% compute commitment, scalable oversight, generalisation, interpretability, adversarial testing and a high-confidence evidentiary standard. **Sutskever, Ilya. “The exciting, perilous journey toward AGI.” TEDAI 2023, October 2023; published 20 November 2023.** Direct: [https://www.ted.com/talks/ilya_sutskever_the_exciting_perilous_journey_toward_agi](https://www.ted.com/talks/ilya_sutskever_the_exciting_perilous_journey_toward_agi) Primary public talk on superhuman intelligence, its transformative potential and his expectation that extreme stakes could produce unprecedented cooperation. **Safe Superintelligence Inc. Launch statement. 19 June 2024.** Direct: [https://ssi.inc/](https://ssi.inc/) Institutional rather than individually authored evidence. Central for understanding the company model Sutskever helped establish: one goal/product, safety and capability developed together, insulation from short-term commercial pressures and “scale in peace”. It should be attributed to SSI unless separately corroborated by him. **Reuters. “Ilya Sutskever on how AI will change and his new startup Safe Superintelligence.” 5 September 2024.** Direct: [https://www.reuters.com/technology/artificial-intelligence/ilya-sutskever-how-ai-will-change-his-new-startup-safe-superintelligence-2024-09-05/](https://www.reuters.com/technology/artificial-intelligence/ilya-sutskever-how-ai-will-change-his-new-startup-safe-superintelligence-2024-09-05/) Secondary publication containing direct interview answers. Crucial personal evidence on safety-gated release, the first SSI product, the limits of the existing scaling formula, and withholding SSI’s primary research while potentially sharing safety work. **Safe Superintelligence Inc. “Updates.” 3 July 2025 entry signed “Ilya”.** Direct: [https://ssi.inc/updates](https://ssi.inc/updates) Official first-person source confirming Sutskever’s CEO role and continued commitment to building safe superintelligence. **Patel, Dwarkesh. “Ilya Sutskever — We’re moving from the age of scaling to the age of research.” 25 November 2025.** Direct: [https://www.dwarkesh.com/p/ilya-sutskever-2](https://www.dwarkesh.com/p/ilya-sutskever-2) The most important late-period primary transcript. It contains the 5–20-year forecast; “age of research” argument; generalisation thesis; explicit change towards incremental deployment; discussion of sentient-life alignment, multiple powerful AIs, power caps, safety collaboration and long-run human political agency. **Safe Superintelligence Inc. and Nvidia. Strategic-partnership announcement, 27 July 2026.** Direct: [https://www.globenewswire.com/news-release/2026/07/27/3333561/0/en/ilya-sutskever-s-safe-superintelligence-inc-and-nvidia-announce-long-term-strategic-partnership.html](https://www.globenewswire.com/news-release/2026/07/27/3333561/0/en/ilya-sutskever-s-safe-superintelligence-inc-and-nvidia-announce-long-term-strategic-partnership.html) Contains a direct Sutskever statement that SSI had research “worthy of scaling up”; important corrective to the interpretation that his 2025 critique of the scaling era implied opposition to massive compute. **Sutskever, Ilya. X post on neocloud cybersecurity and rogue agents. 1 September 2026.** Direct: [https://x.com/ilyasut/status/2094881278621253755](https://x.com/ilyasut/status/2094881278621253755) Latest substantive safety-relevant personal statement located in this review through September 2026. It links autonomous-agent risk to compute infrastructure and calls for stronger cybersecurity. Its characterisation of agents having “gone rogue” should be treated as Sutskever’s assertion rather than silently converted into an independently established fact.