Policy & Regulation Auto-verified translation
Concept
auto-amélioration récursive
auto-amélioration récursive — Concept. definition: Ability of an AI system to direct compute toward its own improvement. Zuckerberg sees this as an inevitable competitive dilemma — any lab that abstains falls behind — and proposes as a safeguard that a "significant majority" of compute remain directed by people, with no numerical threshold, no designated verifier, and no penalty specified · status: Taking hold across the industry since summer 2026, Anthropic included; a candidate for a SALT-style speed cap
Dario Amodei attaches a date to auto-amélioration récursive: since the summer of 2026, AI builds the next generation of AI, across the frontier industry and inside Anthropic. In his September 2026 essay "We Must Pace the Frontier", that is one of two observations that made him revise his position and argue for slowing the rate of capability improvement. The stated risk is a race between two curves: a system pointing compute at its own improvement, against the ability of people to understand and control what results.
Mark Zuckerberg treats the same capacity as a competitive dilemma no laboratory can decline, since a frontier lab that does not direct a substantial share of its compute at self-improvement falls behind intrinsically. His proposed guardrail is that a "significant majority" of compute stay directed by people. The phrase carries no figure, names no verifier, and attaches no sanction.
Amodei's version of the brake is institutional. On his four-level ladder of coordination with China, between an international testing body and an outright general pause, sits a speed cap on self-improvement compared to the SALT treaties.
Eric Schmidt, interviewed by Nicolai Tangen on 21 October 2026, marks discovery as what AI still cannot do: a model fed the mathematics of 1902 does not recover general relativity. He puts AGI five to ten years away. The capacity credited with unlocking discovery is the one Amodei wants paced.
- Type
- Concept
- definition
- Ability of an AI system to direct compute toward its own improvement. Zuckerberg sees this as an inevitable competitive dilemma — any lab that abstains falls behind — and proposes as a safeguard that a "significant majority" of compute remain directed by people, with no numerical threshold, no designated verifier, and no penalty specified
- status
- Taking hold across the industry since summer 2026, Anthropic included; a candidate for a SALT-style speed cap
- relations
- 4
- Cited in
- 2 fiches
Neighborhood
Cited in (2)
Philosophy & Society Auto-verified translation