Ramith Hettiarachchi

if you are doing an A/B test to measure the data mix’s effect on training, should you reset the optimizer states?

Is there a optimized way to run AF3 on tpus? (Comparable to gpu speeds)

This is so good youtu.be/_HZM0QiuUS8?...

did a group presentation today with manim style transitions. Sharing the template for anyone to build upon Brought to life with manim-web + reveal.js + claude + codex view template: slides.ramith.fyi view repo: github.com/ramithuh/pro...

every once in a while opus 5 gets angry with me for not listening to it "Separately — I’d really like an answer on committing." ok chill

For some reason the word tpu “chip” makes me feel like its tiny compared to nvidia gpus 😆 Been playing with v6e chips lately and pretty impressed

The feature where one claude chat can msg other branched off chats is pretty nice..

www.youtube.com/shorts/nJq6x...

astra >> opus 5 (for biology). as a chat goes, Opus 5 keeps on saying > "Your framing was exactly right, and sharper than mine. Let me correct what I said." > "This corrects something I told you earlier."

Does anyone know how to make a NeurIPS rejected paper public? 🧐 is there a deadline to opt in?

ramith.fyi/posts/2026-1...

www.youtube.com/watch?v=WdIf...

Being gpu poor can be a good thing.. u come up with way efficient ways to train models…

wwPDB Remediation of Metal-containing PDB Entries and CCDs Improved accuracy and annotation metal-containing PDB Entries and CCDs are now available in remediated PDB data.

need a quick web tool to understand protein - ligand interactions? try this out pli.ramith.fyi it runs ProLIF in the browser and visualizes the interactions if you want to explore plinder specifically try: --> plinder.ramith.fyi

interesting looking one alphafold.ebi.ac.uk/entry/AF-A0A...

ChatGPT voice mode is really good. Was having several convos about some papers in the literature, asked it to clone the repo etc and asked questions..

Living in the present moment is such a challenge 😆

took me 4-5 days to recover from jetlag 🤦‍♂️

@wheelerlab.org is talking about his effort to build a resource for protein dynamics (PDB for MD). MDRepo is a repository to store simulation data. This is really needed to push the needle in design, function and more. Please help make it successful by sharing your MD data. #ismb2026 🌐 mdrepo.org

Interesting how ai generated music is easy to identify.. has something to do with the lyrics that sound like an essay and the frequency modulations of the voice🤔

is there a repo where esmfold is optimized for somewhat fast folding?

love scientists who have a decent EQ

Claude generated slides kinda give me the ick.. Slides are meant to be simple, somewhat imperfect and human

Registration line 😅 even after two days of registration.. ml conferences maybe reaching a breaking point 😆 i guess it makes more sense to have more decentralized events

Is there a simple way to add bookmarked posters (icml) to the phone’s calendar 🧐

First impression of SK: everything has a touch of cuteness ^_^

So ESMFold2 uses 3DRoPE to enforce a geometry like pairbias (without the mem complexity of the standard pairbias) in its atom transformer… Wonder if the next stage, token transformer could’ve survived a similar approach (rather than the ln proj of z)

damn Grant is an incredible teacher..

<3

Equivariance is dead! 😢 Or is it? 😈 Genie 3 is out! Our latest protein design model achieves SoTA results for binder design and motif scaffolding, greatly improving on BindCraft and Proteina-Complexa. It does so using all-atom SE(3)-equivariance based on a branched polymer representation👇

claude is good at making posters through html.. feels like doing everything programmically is better 😅 used to love editing stuff on illustrator

La proteina’s autoencoder can be thought of as a inverse folder to some degree 🤔 Wonder if just giving Ca, along with a random partial latent acts as a inverse folder (true we get the all atom structure as well)

Consider submitting your work to the GenBio workshop at ICML 2026! We have an amazing set of speakers whose work spans generative and agentic AI. Deadline: May 1, 2026 genbio-workshop.github.io/2026/

❤️

noticing that backbone design models trained on AFDB 588K, when sampled through ODE results in poor designability (~25%). But through a proper noise schedule in SDE it goes to 90% (as reported in lit.) this felt counterintuitive, wonder if the reason is, AFDB being just 42% designable to begin with

this [1] is nice, easier than having to type the passphrase of a ssh key all the time.. made a simpler guide [2] inspired by it [1] - gist.github.com/arianvp/5f59... [2] - gist.github.com/ramithuh/a7b...