Demis Hassabis: What AGI Still Lacks, Whether Agents Actually Work, and What the Next Scientific Breakthrough Looks LikeDemis Hassabis：AGI 还缺什么，智能体到底行不行，下一个科学突破长什么样

Hassabis puts even odds on whether AGI needs one or two more fundamental breakthroughs beyond current pretraining+RLHF+chain-of-thought — continual learning, long-horizon reasoning, and memory remain unsolved.

He uses chess as a diagnostic to show current models lack self-monitoring in reasoning, and argues agents are still experimental because they can't learn on the job.

The interview's opening move sets the tone for everything after: Hassabis refuses a binary answer on whether current techniques are enough for AGI. Instead of hedging vaguely, he commits to a specific number — 50/50 — and names three concrete unsolved capabilities rather than gesturing at 'more compute.' This matters because it separates him from both accelerationist claims that scaling alone gets there and skeptic claims that the whole paradigm is a dead end. DeepMind, he says, is hedging both bets simultaneously.

His memory critique is the most technically grounded section, drawing directly on his hippocampus research. The key insight is that context-window size is the wrong metric — the real bottleneck is retrieval cost, distinguishing signal from noise in what's already stored. He ties this to DQN's experience replay, showing DeepMind's own history already anticipated the problem two AI generations ago, yet it remains unsolved at LLM scale.

The chess anecdote functions as an empirical falsification of the 'reasoning is basically solved' narrative. It's not a benchmark score but a live behavioral observation: the model articulates the correct judgment (this move is bad) yet fails to act on it. This is a sharper critique than typical hallucination examples because it isolates a specific failure mode — lack of self-monitoring — rather than a knowledge gap.

On agents, Hassabis makes a claim that cuts against Google's own product incentives: running dozens of agents for 40 hours often doesn't justify the input, because agents can't adapt to their deployment context without continual learning. This is a rare instance of a lab executive publicly discounting his own company's flagship product category, which lends the critique credibility it wouldn't have from an outside skeptic.

The Move 37 / 'invent Go' distinction and the Einstein test are the piece's most conceptually ambitious moments. Both reframe creativity not as generating a surprising output within a known problem, but as generating the problem space itself. This raises the bar dramatically above current benchmarks (IMO gold, coding contests) and gives a falsifiable, if extremely demanding, test for genuine scientific discovery.

The closing advice to founders — that AGI may arrive mid-journey of a 10-year deep-tech venture — reframes timeline uncertainty as an architectural design constraint rather than a risk to be waited out. Combined with his 'orchestrator + specialized tools' architecture prediction, this suggests domain-specific systems retain value precisely because general models will call them, not replace them.