Hassabis on Lex runs on one conjecture — whatever nature built through selection, a neural net can learn — and everything else follows: an adversarial AGI test staffed by the Terence Taos of each field, science as the proof of benefit, and a posture held like a method — cautious optimism under real uncertainty, with states, not models, as the risk he names most plainly.

The talk's center of gravity is a conjecture: anything nature produced through selection carries structure a classical network can learn. Protein folding and Go are the existence proofs against combinatorial despair.

“All of these are way more than atoms in the universe.”

watch at 12:14

His AGI bar is adversarial, not vibes: hand the system to the Terence Taos of every field for a month or two — fully general means they surface nothing.

“see if they can find an obvious flaw in the system”

watch at 54:01

The differentiation he claims for DeepMind is science itself — AlphaFold as concrete proof of benefit, Isomorphic as its commercial shadow — and a standing invitation to other labs to collaborate on the big questions.

“that's why projects like AlphaFold are so important to me”

watch at 1:39:08

Between star-flourishing and doom scenarios he rules on posture in one sentence — uncertainty plus stakes equals cautious optimism, held as a method rather than a mood.

“the only rational sensible approach is to proceed with cautious optimism”

watch at 1:59:26

The risk register is stated without flinching: states will point AGI at what states point technologies at, and the mitigation he reaches for is minimal international agreement — US-China at least — on basic standards.

“AGI will be used for things that states use technologies for”

watch at 1:55:45

“it may require some agreement internationally”

watch at 2:02:40

The aside that outruns the headline claims: if energy and scarcity are solved, the question is not jobs but denominations.

“What is even money going to mean?”

watch at 1:42:03