𝓐𝓾𝓫𝓻𝓮𝔂 🥀🪦💀

aub.bsky.social

01 August 2026 16:08

[hands raised before an angry mob] ah, no no, you see, I do not use "AI". I use neural networks. LLMs. machine learning algorithms. these are different things & I'm fairly confident you don't know enough about it to know any better

dan

danabra.mov

01 August 2026 16:07

Codex is WAY better than Claude Code. in particular the “Goals” abstraction in Codex was amazing i had one session (“PM”) whose goal was to advance the conjecture, managing a few other sessions (exploratory, math, Lean verification flow) and setting goals for them they mostly stayed on track

dan

danabra.mov

01 August 2026 16:07

Sol seems WAY better at math it found a bunch of seemingly new things that survive adversarial verification by both. they’re not huge results but seem new and not in the literature Fable by contrast seems to either descend into insta-giving up or kept coming up with ultraslop nonsense

dan

danabra.mov

01 August 2026 15:57

ran out of max credits on both Sol and Fable trying to grind a mathematical theorem and get some novel results conclusions so far (thread):

Freya Holmér

freya.bsky.social

01 August 2026 15:57

numbered weeks (1 to 52) and the scandinavian mile (10 km) are pretty useful things I pretty much never see used outside of sweden