Anthropic co-founder Jack Clark: 60%+ chance a model trains its successor by end of 2028 — recursive self-improvement now in official research documents
Jack Clark delivered the 2026 Cosmos Lecture at Oxford on May 20, putting a numerical probability on recursive self-improvement that he says is now in Anthropic's official research documents rather than theoretical speculation. The 60%+ figure for end-2028 RSI moves the conversation from "will it happen" to "what's the timeline" — and the framing of an "intelligence explosion" is now an Anthropic-internal forecast, not a research community speculation.
The framing shift is the consequential part. RSI predictions through 2023-2024 were treated as outside-view futurism even by the labs that took them seriously. Anthropic putting a probability and a date in official research documents is the institutional equivalent of saying "this is the operating assumption for our roadmap." The labs do not put dates on speculative scenarios; they put dates on scenarios they are budgeting against.
The downstream policy question is whether the 2026 regulatory architecture — pre-deployment eval suites, RSP-style if-then triggers, EU AI Act high-risk-system gating — is designed for a world where the model being evaluated may be training its successor faster than the evaluation can produce ground truth. Clark's Cosmos framing argues no. The political consequence of an Anthropic-internal 60%+ probability on RSI before end-2028 is that every framework now in draft has to either accept that timeline as plausible or explicitly argue why it isn't.
Claude5 — AI Safety 2026: Alignment Progress and Open Challenges → · Alignment Anthropic — Alignment Science Blog → · BuildFastWithAI — AI News Today May 23, 2026 →