Ilya Sutskever (OpenAI Chief Scientist) — Why next-token prediction could surpass human intelligence

The episode opens with a promise and an agenda. Dwarkesh Patel's visit to the OpenAI offices in San Francisco is framed less as a conversation than as a checklist of load-bearing questions: how long until AGI, whether models are being abused by foreign governments, what comes after generative models, what life looks like after AGI, the pragmatics of the Microsoft relationship and the Google competition, and the difficulty of aligning a superhuman system. That framing matters for how the rest of the interview reads, because Sutskever answers every question in the same register: short, hedged, and deliberately refusing to over-claim. When Patel's first question is not a question at all but a compliment with a trap in it — the observation that very few scientists make one defining breakthrough and far fewer make several independent ones across a career — Sutskever's answer is the entire thesis of his methodology in miniature. He declines to explain the talent, and reframes the whole thing as effort rather than insight. The implication is that the field's decisive advantage will belong to whoever works hardest, which is either a humbling statement or a competitive one, and is deliberately left ambiguous.

There are not that many scientists who will make a big breakthrough in their field

Read in context, his deflection is doing real work. If breakthroughs are downstream of raw effort, then no amount of careful reasoning about safety or strategy can substitute for the grind itself, and the metric that matters is who stays at the bench longest. It also pre-empts the later claims about how far the interview's answers can be trusted as forecasts: he is signalling from the first minute that he will not be giving calibrated probabilities. Patel's follow-up — that there might be more desirable uses of the same hours than sitting in a lab — is never seriously engaged. The conversation moves on to the first substantive question, and the interviewer's framing about deployment becomes the real opening argument of the episode.

The first substantive question is a pointed one: GPT has been publicly available for months, so why is there not more visible misuse — why are foreign governments not running propaganda campaigns or running scams against elderly targets at scale? Sutskever's answer contains three distinct hypotheses stacked on top of each other, and the order in which he offers them is itself revealing. He starts with the observation that the abuse may simply not have happened in volume, then immediately concedes it would not surprise him if some of it were happening right now, and only then offers the plausible mechanism: adversaries would reach for open-weight models rather than a hosted API, precisely because the open models can be deployed without leaving traces that a provider could observe. The security question is therefore structurally different from the API question, and Sutskever treats it as the more likely one.

I can certainly imagine they would be taking some of the open source models and trying to use them for that purpose

Pressed on whether this is simply a matter of adversaries not having thought hard enough, he offers two more possibilities and ranks neither: they may not have done it at scale using their own technical infrastructure, or it may be happening already, which he calls annoying. He then answers the follow-up about whether such activity could be detected with an unusually flat affirmation — large-scale tracking is possible, yes, but it requires special operations. That last clause is doing the heavy lifting. It concedes that detection is not a routine data-analysis problem but an intelligence problem, which means the cost of concealment scales far below the cost of detection, and the equilibrium is persistent low-level misuse that nobody can fully audit. He offers no discussion of what such tracking would require institutionally, and it is worth noting that he treats the possibility of a foreign adversary extracting model weights and reverse-engineering capabilities as a problem for the entire field rather than a problem specific to OpenAI, which is the more strategically serious of the two concerns Patel raised but the one he addresses more briefly.

From there the interview turns to the economics, and the central puzzle Patel poses is one of timing rather than magnitude. AI is already enormously valuable, but the question is how long the current category of value persists before AGI arrives and collapses the market for human-labour-like work. Sutskever's answer is that the question is partly unanswerable because it depends on where you draw the definition of AGI — which he illustrates with the deliberately deflationary observation that an average college undergraduate already qualifies — and partly because of a well-known bias: people who work on a technology optimistically tend to underestimate the time required to reach it. His corrective is a thought experiment about self-driving cars, in which the system appears to do everything while reliability remains visibly short of solved.

we will need to do some more work until we really iron out all the issues and make it really good and really reliable and robust and well behaved

The subtlety in his answer is the reframe of hindsight. He argues that the window may feel short retrospectively only because each year is larger than the last, and that this creates a systematic bias toward believing the window is closing when in fact the compounding means AI's economic contribution grows every year until AGI arrives. So the window is not narrowing; it is expanding in absolute terms while also being finite in duration. That is a genuinely non-obvious argument, and it is the sort of thing a defender of the industry would want to say and a sceptic would want to test.

When Patel asks for a share of GDP by 2030, Sutskever's answer is the single most methodologically revealing moment in the interview. Asked for an over-under, he declines and explains why with a phrase that ought to be quoted more often than it is:

my error bars are in log scale

An answer in log scale is not a refusal, it is a claim about the shape of the uncertainty: the plausible range for AI's economic contribution in 2030 spans orders of magnitude, and any point estimate would be false precision. The interesting question is what he says next, because he then grants the counterfactual premise Patel constructed — that the value created turns out to be disappointingly small — and identifies the cause. It is reliability. Not capability, not adoption, not regulation, but the mundane engineering property of outputs being right often enough to be trusted without supervision.

That just really puts a damper on the economic value that can be produced by those systems

It is a well-chosen answer. Reliability is the failure mode that scales worst, because it degrades the value of every other improvement simultaneously: a model that is right ninety per cent of the time still requires a human to check every answer, and the checking cost does not shrink with the model. The argument's soft spot is that Sutskever treats reliability as a purely technical problem rather than as one entangled with verification infrastructure and liability. Note also that the interviewer's framing contains an implicit premise Sutskever does not accept: that a technically mature system with imperfect reliability is somehow distinct from an immature one. He rejects the distinction outright — that the systems will be technologically mature, it is only a question of whether they will be reliable enough — and Sutskever's response collapses it.

The episode's central technical question is the one embedded in the title: can next-token prediction actually surpass human intelligence, or is it fundamentally a mimicry engine that can only ever reach parity? Sutskever's position is that the sceptical reading is an artifact of taking the surface description of the method too literally. The standard objection runs: prediction is imitation, imitation cannot exceed its exemplar, therefore next-token prediction caps out at human level. His counter-argument is a shift in what the model is being asked. Instead of asking the network what an average person would do, ask it what a person of unusual insight and capability would do — a person who may not exist, but whose behaviour is nonetheless interpolable from the behaviour of the people who do exist.

I challenge the claim that next-token prediction cannot surpass human performance