Amodei's five hours reduce to one empirical stance: the scaling curve has answered every theoretical objection since 2014, and if it holds, powerful AI lands around 2026-27 — so Anthropic's bet is to industrialize safety on the same schedule: tripwires before capability, incentives over secrecy, with compressed decades of biology as the prize for getting it right.

The capability claim is a slope, not a vibe: real software-engineering tasks went from single digits to half-solved inside one year — the number Anthropic's veteran engineers felt before they believed it.

“it gets something like 50% on SWE bench”

watch at 19:46

“the state of the art was three or 4%”

watch at 19:56

Pressed to a date, he gives one with the error bars showing — powerful AI, his term, plausibly by 2026 or 2027 — and pairs it with the claim that human level is a milestone, not a wall.

“it does make you think that we'll get there by 2026 or 2027.”

watch at 2:19:44

“there's no ceiling at the human level”

watch at 2:06:32

The Machines of Loving Grace half of the conversation: the prize is compressed decades of biomedical progress, gene editing scaled from single edits to whole systems.

“You can do CRISPR, but you can do it for your whole body.”

watch at 2:24:48

Safety operationalized as tripwires, not vibes: ASL thresholds trigger concrete controls before the capability arrives — security against model theft, filters, third-party testing — on a schedule set by the curve itself.

“designed to be sufficient to prevent theft of the model by nonstate actors”

watch at 1:02:55

Anthropic's strategy in one mechanism: publish the safety work so responsibility becomes the industry's baseline — the race to the top reroutes competition itself.

“align the incentives of the industry”

watch at 1:38:15