Twitxy Saved

Janosh

@jrib_

ML for materials discovery @PeriodicLabs. Prev @ Materials Project (LBNL).
SF github.com/janosh Joined October 2020
340 Following 944 Followers 845 Posts

Latest posts 5

Janosh @jrib_
yay! long overdue!
VoidZero Verified account @voidzerodev Oct 20
Introducing tsrs: a Rust port of the TypeScript compiler 🦀

We are not forking TypeScript!

◆ In close collaborating with the TypeScript team
◆ Up to 8x faster than TS 7.0 & ~70% less memory
◆ Runs in WASM, so it works in your browser & on Workers
◆ Soon part of the Oxc toolchain

npx tsrs

blog.cloudflare.com/typescript-rus…
View quoted post on X
Janosh @jrib_
@bcherny @AnthropicAI feature request for Claude desktop: allow fixing the order of projects/envs in the sidebar (drag to reorder, or just alphabetical). they reshuffle whenever an agent finishes, so the one I’m about to click moves last second. gets more frequent/annoying the more agents running
Janosh @jrib_
my thoughts exactly! there's been an overallocation of attention (pun/ambiguity intended) to MLIP architecture development. even though a sheer infinite number of higher impact projects you could do with the architectures we already developed is staring us all in the face!
and if you really want to work on model architecture (rather than generating data in underserved regions of material space or extracting value from existing checkpoints by using them for actual materials science), go add new physical degrees of freedom (charges, spins, ...) to the existing architectures. that expands the science you can do with MLIPs way more than saturating a stale benchmark...
1 replies 4 reposts 40 likes 5.3K views View on X
Janosh @jrib_
VibeMatSci: another hilarious yet educational video by @jainpapers on how well different frontier LLMs do at setting up atomistic simulation cells. tl,dr: not that great yet (even those specifically tailored to the task). small/simple cells are fine but LLMs struggle with larger cells + surfaces/interfaces + anything with compositional/structural entropy... makes only limited difference for now whether the structure is written out directly after some internal reasoning or whether it's the result of ASE/pymatgen tool calling (latter seems slightly better). either way, definitely not good enough to take human experts out of the loop. but the rate of past improvement suggests it could be very different a year or even 6 months from now

youtu.be/FoO820kl2GY

@karpathy might appreciate the video title :)