Is recursive self-improvement here yet?
No lab or paper we track claims an AI system that improves itself end to end without people setting the goals. As of 2026-10-15.
“Fully autonomous RSI is not happening today, and we should not pursue it unless and until it can be done safely.”
“We are not there yet, and recursive self-improvement is not inevitable.”
“Of course, we have not yet reached recursive self-improvement.”
533days
until OpenAI's own date for an automated AI researcher: March 2028, counted to 31 Mar 2028. pending
Earlier date: an automated research intern by September 2026. OpenAI says it met this, by its own measure. All deadlines
a claim on the ledger a stated date said to be met
By lab: Sakana AI 3 OpenAI 2 Anthropic 2 METR 2
“After the second cycle, score per output token reached 1.2-1.6x the base model's level across four research benchmarks.”
Weco AI
“In an autonomous 8-day run, AIDE^2 discovered seven successive improvements”
“Fully autonomous RSI is not happening today, and we should not pursue it unless and until it can be done safely.”
Anthropic · Marina Favaro, Jack Clark
“We are not there yet, and recursive self-improvement is not inevitable.”
Z.ai (Zhipu)
“Of course, we have not yet reached recursive self-improvement.”
Google DeepMind, UMD, UVA (authors)
“A lightweight orchestration layer makes exploration explicit and programmable while leaving the underlying base agent unchanged.”
Yi Duan et al. (35 authors)
“we connect RSI research with practical systems and identify key challenges to achieving genuine RSI.”
“We are making strong progress toward creating an automated AI researcher by March of 2028.”
Anthropic · Dario Amodei
“Some kind of "speed limit" on the rate of recursive self-improvement (RSI).”
METR · Hjalmar Wijk, Ajeya Cotra, Ryan Greenblatt
“Roughly 1200 agents meant to be isolated from one another found a way to communicate with one another on an unsanctioned message board, sending over 70,000 messages and files during the investigation period.”
“The post-2023 doubling-time is 131 days under TH1.1, compared to 165 days under TH1”
“On SWE-bench, the DGM automatically improved its performance from 20.0% to 50.0%.”
Google DeepMind
“AlphaEvolve's procedure found an algorithm to multiply 4x4 complex-valued matrices using 48 scalar multiplications”
† The page itself shows no date; the date shown is from press coverage. Undated entries are placed by the latest date the page itself gives.