Free tools Windows power users keep installed
One-click scans. No signup required.
Jack Clark, Anthropic’s co-founder and author of the Import AI newsletter, now says it is more likely that “full recursive self-improvement” (RSI) will become possible in 2027 rather than 2028. The statement is Clark’s personal forecast in Import AI issue 476, dated October 12, 2026—not an Anthropic product announcement, official company target or proof that full RSI has arrived.
What Clark changed
Clark wrote that his estimate had moved forward by one year: “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.” He presented the change as a probability judgment about when the capability might become possible, not as a scheduled milestone.
As an Amazon Associate I earn from qualifying purchases.
His explanation was that AI systems have made substantial advances in creative-seeming mathematical breakthroughs and are showing stronger performance on tasks related to AI research. In Clark’s account, creativity had been a bottleneck. New observations about mathematical creativity and research performance led him to revise the estimate.
What “full RSI” means
Full recursive self-improvement describes a feedback loop in which an AI system contributes to improving the systems that will replace or succeed it, with those successors then able to improve again. The important feature is the recursive loop, not simply that software can perform useful research work.
#1 Best Overall
An AI that writes experiment code, searches scientific literature, proposes model changes or automates parts of an engineering team’s workflow may accelerate research without demonstrating full RSI. Research assistance and research-labor automation are therefore weaker claims than an autonomous or largely autonomous process that produces materially more capable successor systems.
Why the TasteVal figures matter to the forecast
Clark cited benchmark trends as part of the context for his judgment. Import AI issue 476 reproduced a statement from TasteVal saying that the “experimental research taste” of frontier models had doubled every 3.0 months since December 2025, with a reported 95% confidence interval of 1.7–5.0 months.
Rank #2
TasteVal also reported that its best model, identified in the statement as Opus 5.5, exceeded the benchmark’s human baseline with a 2.30x compute multiplier. The reported 95% interval was 1.15–4.37x, and the model’s cost per attempt was described as roughly one-thirtieth of the expert baseliners’ average cost per attempt.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Scan for outdated or missing drivers - takes under a minuteDriver Scan →| Reported measure | Figure | Qualification |
|---|---|---|
| Rate of benchmark improvement | Doubling every 3.0 months | TasteVal estimate reproduced in Import AI; 95% confidence interval: 1.7–5.0 months |
| Best-model compute multiplier | 2.30x | TasteVal result as reported by the newsletter; 95% interval: 1.15–4.37x |
| Cost per attempt | Roughly 1/30 of expert-baseliner average | TasteVal description reproduced in issue 476 |
| Human baseline | 24 experts | At least two experts per task, according to the newsletter’s account |
The reported human baseline comprised 24 experts who had recently worked at organizations including OpenAI, Google DeepMind, NVIDIA and Microsoft Research. The newsletter also described slower improvement in earlier years and a change in slope around late 2025 or early 2026.
Rank #3
These numbers are evidence Clark considered relevant, not a demonstration that full RSI will occur in 2027. The underlying benchmark methodology and primary report were not independently examined for this article, so the figures should be read as TasteVal’s reported results rather than as independently verified proof of a recursive improvement loop.
Is Anthropic predicting RSI in 2027?
No. The 2027 statement comes from Clark’s newsletter and is framed as his own updated view. Anthropic has not announced a 2027 RSI target, and the statement does not establish that the company has a system capable of full recursive improvement.
A secondary analysis of terminology also distinguishes Anthropic roadmap language about automating research from a specified recursive mechanism. Planning for safeguards around possible research automation is not the same as announcing that an AI can autonomously improve its successors.
How this compares with other public milestones
Associated Press reporting in September 2026 said Anthropic had not expressly stated how close it was to fully autonomous model improvement. The same report described OpenAI’s goal of an automated AI researcher by March 2028. OpenAI also cautioned that rapid RSI is not necessarily an outcome to pursue and that safe progress is uncertain.
Best Value
Those statements provide context but do not confirm Clark’s forecast. They concern different organizations, dates and capability definitions. An automated researcher could be a major step toward RSI while still falling short of a closed, self-improving system.
What would count as evidence of full RSI?
- A defined improvement loop: the system would need to help create or select successor systems, not merely perform isolated research tasks.
- Measured capability gains: successors would have to demonstrate meaningful improvements attributable to the loop.
- Operational autonomy: evidence would need to show how much human intervention, approval and infrastructure management remain necessary.
- Reproducible evaluation: claims should specify the benchmark, baseline, compute budget, costs and uncertainty intervals.
Until those details are available, benchmark gains in mathematical creativity or AI-research skill are best treated as indicators of progress toward more capable research systems—not confirmation of full RSI.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
