On the record
“AI systems haven’t yet displayed this [creativity] in a transformative and major way (though some of the results on accelerating math research are suggestive of this).”
“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.”
Opinions and predictions (not established facts)
- Full recursive self-improvement (RSI) is more likely to become possible in 2027.
- AI systems are developing better research taste, enabling faster self-improvement.
What changed
In the latest Import AI newsletter, Jack Clark discusses the concept of recursive self-improvement (RSI) in AI systems. He notes that recent benchmarks, such as TasteVal, indicate that AI models are developing better "research taste" – the ability to identify interesting problems, design experiments, and interpret results. Clark argues that this improvement in creative and research-oriented capabilities makes RSI more likely to occur sooner than previously thought. He revises his prediction for the advent of full RSI from 2028 to 2027, citing advancements in AI's capacity for creative mathematical breakthroughs and AI research itself.
Why it matters for builders
Clark's revised timeline for RSI suggests a potentially faster trajectory for AI development. If AI systems can effectively improve themselves, the pace of innovation could dramatically increase, presenting both opportunities and challenges for AI builders and researchers. Understanding this accelerated timeline is crucial for strategic planning and development in the AI field.
Practical impact
AI on Radar's take: Clark's updated RSI timeline is a significant shift. The notion that AI can increasingly perform research tasks and potentially self-improve at an accelerated rate warrants close attention. Builders should consider how this might impact the competitive landscape and the types of AI capabilities that will emerge in the near future. The increasing sophistication of AI in research tasks, as highlighted by benchmarks like TasteVal, suggests a move towards more autonomous AI development cycles.
Caveats and source limits
The article presents Clark's analysis and predictions regarding AI's recursive self-improvement. The timeline for RSI is a prediction and not a verifiable fact. The TasteVal benchmark measures aspects of research taste, but its direct correlation to full RSI is an interpretation by Clark. The source does not provide independent verification of the benchmark results or the predicted RSI timeline.
Sources
Claim check: 3/3 supported claims - 3 evidence links - 88% avg confidence
- AI systems are developing more of the creative intuitions which human researchers ascribe to ‘good scientific taste’.supported - importai.substack.com
- The experimental research taste of frontier models has doubled every 3.0 months since December 2025.supported - importai.substack.com
- Full recursive self-improvement (RSI) is more likely to become possible in 2027, revised from a previous prediction of 2028.supported - importai.substack.com
Caveats
- This is based on the TasteVal benchmark and Jack Clark's interpretation.
- This is a finding from the TasteVal benchmark as reported by Jack Clark.
- This is Jack Clark's updated prediction based on current trends.
- Single-source caution: verify critical details at the linked source.
Radar score 77/100 - how it was calculated
- Reliability 77: Primary on-record statement
- Freshness 100: Fresh source date
- Novelty 64: New statement by a tracked AI voice
- Technical 56: Structured technical source signals
- Developer 50: Builder relevance source signals
- Ecosystem 70: Tracked AI voice
- Confidence 98: Claims have reliable evidence
Discussion
Loading comments...