ok i take it back Astra is good how to work with it: 1. get it to deeply understand what needs to be built. not “plan“ but like get it to be a domain nerd 2. only THEN, give it a usable past sloppy attempt and ask for excellence 3. ask it to use Sols for coding so it stays on strategy and taste
re last point, can somebody from frontier labs please finally drill polya’s “how to solve it” into the models thinking trace habits? “we have to shift our position again and again“ is so obvious to how humans work but models refuse to do that because they race towards the finish line! so annoying
5.6 Sol is my guy. best model in town right now imo. maybe not the cleverest but at least it’s relatively dependable. especially if you give it a little bit of structure on how to work
wish i could refund 95% of my Astra usage. keep giving it a chance and it fumbles but it spends tokens way faster than Sol. a very disappointing release for actual implementation work
thanks to Dr. Samuel Allen Alexander for making a video about my post! www.youtube.com/watch?v=CFkr...
Conway's Conjecture AI-proved, with AMAZING writeup!
YouTube video by Dr. Samuel Allen Alexander
youtube.com
i spent all of my free time for a month on this. it's done. so i wrote about it.
my uninformed mental model of AI proofs in mathematics is that they’re like lighthouses in the fog. the fog is still there, and clearing the fog is the primary value of the discipline. the lighthouses give a bit of an orientation but don’t clear the fog on their own.
annoying bsky app regression: pressing profile icon (or profile posts tab) on web no longer invalidates it
Frog built a wet lab for the AI model. "There," he said. "Now it can do its own experiments." "What the fuck?" said Toad
another fantastic essay on the current thing
Last week I posted my essay on LLMs, cultural technologies, and mathematics. It is my earnest attempt at convincing mathematicians that should be optimistic, even prideful, at this critical moment: birdsnfrogs.github.io/2026/09/12/F... 1/
Last week I posted my essay on LLMs, cultural technologies, and mathematics. It is my earnest attempt at convincing mathematicians that should be optimistic, even prideful, at this critical moment: birdsnfrogs.github.io/2026/09/12/F... 1/
Happy, those able to know the causes of things
birdsnfrogs.github.io
great guest post by @3blue1brown.com
If math is more than proof, we need to better celebrate the rest of it
[This is a guest post by Grant Sanderson. This blog post was initially written in a different file format and converted using AI. — T.] A sentiment echoing throughout the mathematics communit…
terrytao.wordpress.com
anyone hooked up Jev to any proof related stuff? can it be useful for Lean? i haven't learned much about it yet
i spent all of my free time for a month on this. it's done. so i wrote about it.
How I Vibed a Proof of Conway’s Conjecture — overreacted
You can just prove things, apparently.
overreacted.io
Increasingly convinced that specs are now the source and that code is an artifact akin to a lockfile.
🙋♂️ ever wanted to build something without needing to configure a db, auth, file storage or a cms? it turns out you already have all four if you have a bluesky account 👉 airspace is a typed client for it
airspace
The database you already have.
getair.space
i want something like @rescript-lang.org but for @lean-lang.org. that is, a JavaScript backend emitting reasonably idiomatic JS code. i wonder whether this is structurally impossible.
fractional calculus is so cool! if i understand correctly it's like when instead of a first derivative (f') or second derivative (f'') or say antiderivative (∫f dx), you can take ... half derivative
4/ Standard proteome allocation models (Markovian, one timescale) cannot reproduce this. They predict cells should fully reset between pulses. Real cells don't. So we rewrote the dynamics using fractional calculus, which has memory built in. The new model quantitatively explains the data.
learning this over and over and over. really best way to go with frontier models
i think the main thing i learned with my llm math detour (and now see with speculative llm projects in general) is that telling llm to burn it all down and start from scratch when you learn an important lesson (or ask it what it wants to salvage) brings as much, if not more, value than construction
proving a relatively nontrivial thing: consistently bad output, going offtrack, faking progress. i had to intervene and think with it through every step very closely for entire day finally, proof compiles. "astra, find ways to simplify it". immediately, many big reductions, proof is much simpler
so far i find its greatest strength is looking at a shitton of code and seeing patterns / things to simplify / better approaches. but it's not gonna do that proactively to its own work, and it's too clever for its own good
understated but both of these are kinda huge. ViewTransition is the first ever (!) first-class animation API in React. (it's powered by the browser API but is composable in a very Reacty way.) and Fragment refs solve "merge refs" soup. you can now have natural apis like <IntersectionObserver> etc
React 19.3 is now available! This release makes View Transitions and Fragment Refs stable, and adds browser(), Trusted Types support, and Context in Server Components. react.dev/blog/2026/09...