Anthropic cofounder Jack Clark moves his forecast for full recursive self improvement to 2027 from 2028
Jack Clark of Anthropic now thinks it's more likely we see full RSI become possible in 2027 than in 2028.
Context
Clark helps shape Anthropic's policy views, and an insider pulling forward the date AI could improve itself raises the stakes for safety and regulation debates.
Rob Wiblin hosts the 80,000 Hours podcast, but the forecast comes from Clark's own newsletter, Import AI issue 476, published October 12. This is Clark's personal forecast, not an Anthropic announcement. In May he put roughly 60% odds on automated AI research by the end of 2028 and about 30% on 2027, according to coverage of Import AI 455.
Deep dive
On October 13, Rob Wiblin posted: Jack Clark of Anthropic now thinks it's more likely we see full RSI become possible in 2027 than in 2028, with a screenshot of Clark's newsletter.
In Import AI 476, Clark wrote: I now think it's more likely we'll see full RSI become possible in 2027, whereas before I thought it was more likely in 2028. RSI means recursive self improvement, a point where AI systems can do the work of building better AI systems.
Clark had earlier named creativity as the key missing ingredient. He wrote that AI has since made creative seeming math breakthroughs, and pointed to TasteVal, a new benchmark from P Zero Research that tries to measure research taste. AT Pulse reported that TasteVal found frontier models' experimental taste doubling about every three months since December 2025, and that Opus 5.5 beat the best human expert attempts on eight AI research tasks at about one thirtieth of the cost.
Think Facility noted that Clark gave no new percentage for 2027, and that he has warned that alignment methods that work today may break under recursive self improvement. TasteVal is one new benchmark, which Clark himself has cautioned against overreading.
What to expect next
Watch whether other lab leaders revise their timelines after Clark's change, and whether Anthropic's policy positions shift to match a 2027 forecast.
Also watch for independent replications of TasteVal, since Clark's change leans on one new benchmark, and for the reaction from critics who say labs talk about pacing while racing, as Derek Thompson argued the same day. Any new Import AI issue that updates his probabilities would show how firm the change is, and whether Anthropic's leadership shares it.