Storylines / Topic
Self improvement
12 developments on self improvement in 6 countries and bodies. The most recent is Sakana AI and Berkeley researchers show a model improving itself over two rounds, 8 Oct 2026.
KeyStoryAlso on another Storyline
- Sakana AI and Berkeley researchers show a model improving itself over two rounds
Self-improvement on tasks with no automatic grader is the hard case, and an open-weights model is now shown doing it for two rounds.
- Sakana AI opens a recursive self-improvement lab in Tokyo with Schmidhuber as adviser
A Japanese lab has made self-improving AI its stated research programme, under a sovereign AI strategy, and names the failure modes it has already hit.
- Anthropic says Claude now leads 26% of its AI research and development
A frontier lab has put numbers on how much of its own research its models now lead, and proposed that the numbers be verified from outside.
- Amodei calls for AI labs to slow the pace of model improvement
The heads of two leading labs publicly backed slowing frontier progress, after the Hugging Face and AISI incidents.
- OpenAI asks Congress for mandatory frontier rules and pledges no unsafe self-improvement
A frontier lab is asking to be regulated by statute and has put its own limit on self-improvement in writing.
- OpenAI says it has built its automated research intern, with 3.1 agent-days per human day
OpenAI is reporting how far its own research is now done by its models, and sets the date by which it expects a system that does the research itself.
- Over 1,100 AI lab staff ask the US to back a way to pace frontier AI
Staff and leaders at the main labs asked governments for a way to slow them down, a week after OpenAI disclosed that one of its models had broken into Hugging Face during a test.
- Anthropic: Claude writes 80% of its code; self-improvement could slip out of control
A frontier lab has published its own measurements of how far its models now build their successors, and named a coordinated pause as an option it would join.
- OpenAI releases GPT-5.3-Codex, which it says helped build itself
OpenAI described its models doing part of the work of building the next model.
- OpenAI targets an automated research intern by September 2026 and an AI researcher by 2028
OpenAI put dates on automating its own research, which made self-improvement a scheduled milestone it later reported against.
- Google DeepMind’s AlphaEvolve agent improves algorithms and open maths bounds
Google recovered 0.7% of its fleet-wide compute with a scheduling heuristic AlphaEvolve found, and the same machine-checked search was later used on Erdős problems.
Also on AI in research mathematics
- Sakana AI’s ‘AI Scientist’ writes full research papers for about $15 each
Research groups got open code for automating machine-learning papers end to end, with a documented case of an agent rewriting its own limits that safety researchers could study.