Congrats @rikiyatakehi and team, this is the most exciting result I've seen in a while! Two ways I'm reading this:
- Weight optimization is a text optimization problem once your optimizer is good enough (it's just a long structured list of floats)
- Meta-learning is finally
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming:
I think there’s a specific type of thinking that LLMs will be bad at for a while. You might call it “conceptual engineering”: designing abstractions from first principles while letting experiments bend your axioms.
Related to "applied pure mathematics" argmin.net/p/applied-pure…
it’s my nature to keep posts rather impersonal, so i have to work extra hard to tweet this sort of personal testimony below from time to time as i think the reminder in it around AI and the “value of human intelligence” is worth it
as of the end of Oct 2026, i have not (yet -
ScholarCatalyst is part of Slingshots Four!
Deeply “reading” the literature, i.e. knowing which earlier idea a new problem needs, takes years to get even okay at. Even the best systems do poorly on ScholarCatalyst. With @LaudeInstitute's support, we’ll train directly for this.
ScholarCatalyst // @yoonholeee, @ohmyksh, @chelseabfinn (@Stanford) — A benchmark for finding prior work that could inspire scientific research, and a retrieval system trained for this task.
⏰ Happening TOMORROW!
The 2026 #MichiganAI Symposium brings together researchers, industry leaders & students to explore foundational models & their impact on academia, industry and society.
🎤 Keynotes, talks, posters & more, join us at @UMichCSE!
ai.engin.umich.edu/events/2026-ai…
