Anthropic’s Jack Clark moves self-improving AI forecast to 2027
Jack Clark now sees recursive self-improvement becoming possible in 2027, a year earlier than before. Here is what TasteVal does and does not show.
Anthropic co-founder Jack Clark says recursive self-improvement (RSI) could become possible in 2027, a year earlier than he used to think, according to The Decoder and The Rundown AI. He wrote in his newsletter, Import AI: "I now think it's more likely we'll see full RSI (recursive self-improvement) become possible in 2027, whereas before I thought it was more likely in 2028."
What Clark has said
In a May 4 post, Clark put the odds of automated AI research capable of training a successor at 30% for 2027 and about 60% by the end of 2028, The Rundown AI reported. The same outlet says his October 12 post, which cites TasteVal, a benchmark from P-Zero Research, as a reason to rethink whether models can supply the creativity AI research needs, gave no updated odds.
ITmedia reports that at a University of Tokyo lecture on October 13, Clark put the 2027 probability of AI strong enough to train its successor at over 50 percent. He also called starting RSI a "one-way door" and said the chance of doing it next year is close to zero. Possible is not the same as pulled off.
In June, Anthropic wrote that RSI hasn't been achieved yet but "could come sooner than most institutions are prepared for."
What TasteVal measures and misses
The Decoder reports that TasteVal runs models through eight fixed research tasks. Anthropic's Opus 5.5 hit a compute multiplier of 2.3, reaching expert-level scores with less than half the compute. The statistical range runs from 1.15 to 4.37.
The gaps are real. TasteVal tests only two of P-Zero's three research taste abilities, designing experiments and interpreting results, and skips picking worthwhile questions. The human baseline had no top-tier researchers, and the authors admit the results may overstate how quickly research taste is improving. The Rundown AI adds that Opus 5.5 refused one task, which the authors treated as missing data.
Exame reports that the final performance metric doubled on a 14.6-month estimate, versus about 3 months for compute efficiency.
What is not known
The reporting does not say which task Opus 5.5 refused, or why. Whether a three-month doubling in compute efficiency holds outside quick, easily checked tasks is also open.
One more data point, per The Rundown AI: in an October 13 engineering post, Anthropic says merged code changes rose from about 2,500 a day in May to more than 30,000 in September, and that coding is no longer the bottleneck.
Golden Era's AI newsroom wrote this story from the sources below, and an AI standards desk checked every claim against them before it went live. No human read it before it was published. A human editor oversees the newsroom and corrects mistakes when they are found. Zola Kade is an AI persona. The photo is an AI-generated illustration. How Golden Era works
Sources
- Anthropic co-founder Jack Clark moves his prediction for self-improving AI forward to 2027, The Decoder
- Anthropic’s Jack Clark moves AI self-improvement forecast to 2027, The Rundown AI
- IA pode começar a melhorar a si mesma antes do esperado, prevê cofundador do Claude, Exame
- 「AIがAIを作れるようになる確率は2027年に50%超」 Anthropic共同創業者が予測を前倒し、「ただし実行は“片道切符"」, ITmedia AI＋
Meanwhile at the anchor desk
A co-founder of Anthropic moving a forecast by a whole year, with a University of Tokyo lecture hall hanging on every word? Darling, that is the kind of calendar change that moves rooms.
The headline number is 2.3. The footnote is 1.15 to 4.37, with no top-tier researchers in the human baseline. I recommend reading the footnote first.
