Liam Bai
@liambai21
writing about math, AI, and biology | software @ginkgo, math + CS @brownuniversity
Remember Golden Gate Claude? @etowah0 and I have been working on applying the same mechanistic interpretability techniques to protein language models. We found lots of features and they’re... pretty weird? 🧵
Sartre needs to be rehashed in light of LLMs… many ideas are quite deep and relevant today. Left: Being and Nothingness Right: the void, by @nostalgebraist (a great read) Consciousness is “empty”; nothingness is at the heart of being… LLMs are similar by construction.
nostalgebraist has written a very, very good post about LLMs. if there is one thing you should read to understand the nature of LLMs as of today, it is this. I'll comment on some things they touched on below (not a summary of the post. Just read it.) 🧵 nostalgebraist.tumblr.com/post/785766737…
MechInterp for pLMs has been considered a “dark art” for some time due to spurious correlations and pseudo-insights into biological mechanisms Happy to have contributed a bit to this ongoing effort, containing many stellar, honest results. Plenty of work yet to be done!
Update to our ICML paper on interpreting features in protein language models: We ask human raters to assess the interpretability of SAE feature activations and ESM activations. Human raters found SAE features far more interpretable! x.com/etowah0/status…
Update to our ICML paper on interpreting features in protein language models: We ask human raters to assess the interpretability of SAE feature activations and ESM activations. Human raters found SAE features far more interpretable! x.com/etowah0/status…
Can we learn protein biology from a language model? In new work led by @liambai21 and me, we explore how sparse autoencoders can help us understand biology—going from mechanistic interpretability to mechanistic biology.
You can take a ESM's final-layer projection matrix (used to get logits) and apply it to intermediate layer embeddings. This can roughly tell you what the model “thinks” the sequence is at different layers. x.com/liambai21/stat…
I wrote a new blog post on using the logit lens to interpret protein language models! It has some interactive visualization you can play around with 👇
Proud that this work was accepted at ICML as a spotlight poster! Reminder that you can just do things, especially in interpretability. @etowah0 and I did most of this work in colab notebooks that cost a few dollars a month.
Can we learn protein biology from a language model? In new work led by @liambai21 and me, we explore how sparse autoencoders can help us understand biology—going from mechanistic interpretability to mechanistic biology.
Learning to code will be like learning math: you’ll always use a calculator, but it teaches you how to think
I no longer think you should learn to code.
🚀 Fellowship applications are OPEN for Encode: AI for Science. What if you could use AI to - Design shape-shifting robots - See through solid materials - Decode language of the brain - Create advanced materials
If you’ve ever - thought AI protein folding is magical ✨ - wanted more than a pLDDT score 🔎 - or just think mech interp in bio is cool 🤓 then read the 🧵 👇 on our first paper towards interpretable protein structure prediction just accepted to workshops at ICLR
A First Step Towards Interpretable Protein Structure Prediction With SAEFold, we enable mechanistic interpretability on ESMFold, a protein structure prediction model, for the first time. Watch @NithinParsan demo a case study here w/ links for paper & open-source code 👇
I started @cypherbio last month after 8 years @Ginkgo working on the intersection of software and biology! At Cypher, I am building AI enabled software tools to supercharge scientists productivity and capabilities. Check this out! loom.com/share/2e18e219…