Shane Legg (DeepMind Founder) — 2028 AGI, superhuman alignment, new architectures

Dwarkesh opens with what sounds like a straightforward instrumentation question: loss curves improve from one model to the next, but a loss number is just a number, so how should we read it as progress toward AGI (s003). Legg does not answer the question on its own terms. He moves the problem upstream, arguing that the difficulty is definitional rather than metrical: “AGI by its definition is about generality.” The consequence he draws is that narrow measurement is easy precisely because narrowness permits test construction — “It’s much easier to measure performance when you have a very specific thing in mind because you can construct a test around that” (s004). This is correct, and it also explains why the loss curve cannot be rescued by better curve-reading. A scalar trained on next-token prediction measures competence at a task family, not breadth across task families, and no amount of interpretability applied to the scalar converts one into the other.

The definition he then offers is deliberately modest: “a machine that can do the sorts of cognitive things that people can typically do, possibly more” (s005). Two features of that phrasing matter. First, the threshold is set at typical human performance, not at the best human performance and not at superhuman performance; superhumanity is explicitly optional (“possibly more”), which means the bar is a floor rather than a target. Second, the operative word is “typically,” which quietly imports a statistical notion — the reference class is what people ordinarily manage, not what a specialist manages after training. Legg does not say how the distribution over “typically” is to be estimated, and that omission recurs later when he needs a human baseline to compare against.

What he does supply is a procedure. If the goal is to know whether the threshold has been met, the requirement is a large and varied battery of measurements spanning the breadth of human cognitive tasks, plus a sense of human performance on each (s006). The stopping rule is the interesting part: you have AGI when you have a good range of tests, a system that meets human performance on all of them, and then — “even with effort, you can’t actually come up with new examples of cognitive tasks where the machine is below human performance” (s007). This is not a checklist but a search procedure. Its termination condition is the failure of a search, not the satisfaction of an enumeration.

Legg is explicit that the criterion is epistemic rather than ontological. It may be conceptually possible that some human capability is missing, but “if you can’t find it with some effort, then for all practical purposes, you have an AGI” (s008). The honesty here is worth noting, and so is the cost. On this account, whether a system counts as AGI depends partly on a property of the testing community — its effort, ingenuity, and willingness to look — rather than solely on the system. A sufficiently lazy or under-resourced evaluation effort would certify AGI earlier than a rigorous one, and the definition provides no way to distinguish the two cases from the inside. Legg never specifies who is supposed to expend the effort, how much effort counts as sufficient, or what happens when the search is inconclusive rather than successful.

Pressed for concreteness against MMLU and similar benchmarks, Legg’s first example is streaming video (s010). It is worth noticing that this answers a slightly different question than the one asked. Dwarkesh asked which aspects of human cognition are inadequately measured; Legg’s example is a modality the models do not ingest at all. That is a real and large gap, but it is an input-channel gap rather than a faculty gap, and the two come apart: a system could handle video tokens and still fail at the cognitive work video understanding requires. Legg’s phrasing — “They don’t measure things like understanding streaming video” — keeps “understanding” in the sentence, so the stronger reading is available, but the argument as given rests on the modality.

The second example is sharper. Legg lays out a three-tier picture of human memory: working memory for what happened very recently, cortical memory for what has been consolidated into cortex, and a system in between — episodic memory, implemented by the hippocampus — which he describes as being about “learning specific things very, very rapidly” (s011). The illustration is concrete and personal: if you remember tomorrow what he said today, that is the hippocampal system at work. The claim about current models is then blunt: they do not have this, and there is no test for it (s012). The substitute is architectural brute force — lengthening the context window, which he classifies as closer to working memory, “to sort of compensate for this.” The hedge “sort of” is doing real work in that sentence; it concedes that compensation is approximate.

Asked why this is hard to pin down, Legg gives the most defensible answer in the segment: the generality of human intelligence is very broad, so identifying what is missing requires going “into the weeds” to find specific missing types or entirely absent benchmark categories (s013). He does not claim a complete taxonomy. That restraint is appropriate, but it also means the gap analysis remains a set of examples rather than a framework, which is exactly the thing his definition of AGI would seem to require.

Dwarkesh proposes a reduction: is episodic memory just sample efficiency? Legg accepts the connection immediately — “It’s very much related to sample efficiency” — and adds that it is one of the things that makes humans sample efficient in the first place (s015). He then distinguishes kinds of learning. Large language models do have a form of it: once something is in the context window, the conditional distribution shifts and behaviour changes, which is a very rapid kind of learning. But learning is plural, and “the existing systems have some of them, but not others.”

The trillion-token contrast in the next question is the strongest version of the objection, and Legg’s reply is the most technically specific claim in the first half of the interview. He describes three timescales: immediate learning when something is in context, slow learning during base-model training over trillions of tokens, and nothing in between — “But they miss something in the middle” (s017). This reframes the sample-efficiency worry. The complaint is not that models learn too slowly in general, but that they lack the middle band: the regime where a single exposure produces durable, retrievable, specific knowledge. That band is precisely where hippocampal function sits in his earlier account, which makes the three-tier memory picture and the three-timescale learning picture the same argument stated twice.

He then declines the fatal-flaw framing: “I don’t think it’s a fundamental limitation” (s018). The supporting claim is that something fundamental has already changed — models now have some degree of understanding of what is going on, which did not exist before, and the method is scalable, which unlocks further work. From there he lists the shortcomings he expects to be addressed: delusions, factuality, the type and timescale of memory and learning, video understanding (s019). The conclusion is carefully hedged: no big blockers, no walls, more research, and the problems will “probably be adequately solved.”