It feels less like “self-improvement” and more like fast outer-loop search over scaffolding. We’re basically seeing hill-climbing on prompts, tools, and workflows with tight feedback loops.
Four "labs" shipped self-optimizing AI, two in the last month. MiniMax ran 100+ rounds of autonomous scaffold optimization (30% gain). Karpathy's autoresearch (630 LOC) ran 700 experiments in two days. AlphaEvolve beat Strassen's 1969 matrix multiplication record. Microsoft's STOP did it academically in 2023. None retrained weights. They all optimize the agent layer: prompts, tools, sampling params, workflow logic. Schmidhuber's Gödel Machine needed formal proofs before self-modification. What worked was brute force with a 5-minute timeout.
Open question: what happens when the optimizer can also rewrite its own eval function?
It feels less like “self-improvement” and more like fast outer-loop search over scaffolding. We’re basically seeing hill-climbing on prompts, tools, and workflows with tight feedback loops.
tl;dr
Four "labs" shipped self-optimizing AI, two in the last month. MiniMax ran 100+ rounds of autonomous scaffold optimization (30% gain). Karpathy's autoresearch (630 LOC) ran 700 experiments in two days. AlphaEvolve beat Strassen's 1969 matrix multiplication record. Microsoft's STOP did it academically in 2023. None retrained weights. They all optimize the agent layer: prompts, tools, sampling params, workflow logic. Schmidhuber's Gödel Machine needed formal proofs before self-modification. What worked was brute force with a 5-minute timeout.
Open question: what happens when the optimizer can also rewrite its own eval function?