26 October 2026
Self-improving AI agents show early gains, but not yet true takeoff
First reported
Exponential View ran this on .
- Researchers are testing AI systems that improve themselves, tracing the idea back to the 1960s thinker I.J. Good.
- Recent examples include Sakana AI, Google DeepMind's AlphaEvolve, and MiniMax's M2.7, which got better at tasks.
- The gains came mainly from changing the surrounding software, such as instructions and memory, not from retraining the core model.
How it was covered
Exponential ViewAzeem Azhar
Recursive self-improvement history, Sakana, AlphaEvolve, MiniMax experiments, Sundar Pichai's comments, and Ramez Naam's taxonomy