Jack Clark moves self-improving AI forecast forward to 2027
Original title: Anthropic co-founder Jack Clark moves his prediction for self-improving AI forward to 2027
TasteVal gives Anthropic's Opus 5.5 a 2.3 compute multiplier, but it does not show models already improving themselves.
Anthropic co-founder Jack Clark now says recursive self-improvement could be possible as early as 2027, one year sooner than he previously expected. He cites TasteVal, a P-Zero Research benchmark of experiment choice and result reading, where Opus 5.5 matched expert scores with a 2.3 compute multiplier. The test does not cover choosing research directions, and its authors say it may overstate how fast research taste is improving.
Unconfirmed: Clark's 2027 date is his forecast, and P-Zero says TasteVal may overstate the pace of improvement.
- In May he put the odds at about 30 percent by 2027 and 60 percent by the end of 2028.
- Opus 5.5's reported compute-multiplier range runs from 1.15 to 4.37.
- The stated doubling time for research taste ranges from 1.7 to 5.0 months.
- Each TasteVal run is capped at 40 GPU-hours or 120 wall-clock hours on one H100.
- A Fudan study found humans still made over 85 percent of final research decisions.
- Vendors
- Anthropic
- Organizations
- P-Zero ResearchFudan University
- AI models
- Opus 5.5
From the original article
5 min read at the-decoder.com
A new benchmark by P-Zero Research measures how well AI models design and evaluate experiments on their own. The results are improving so fast that Anthropic co-founder Jack Clark now thinks self-improving AI could arrive by 2027. The article Anthropic co-founder Jack Clark moves his prediction for self-improving AI forward to 2027 appeared first on The Decoder.
By Manuel Uth · the-decoder.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.