🎓 Hats off to the 2025 IICD graduates: Yining Ma Junze Huang Yichi Yang Ruilin Dai Boan Zhu Cameron Park @jlfan.bsky.social & Achille Nazaret!
Wishing you all the best in your next chapter — we’re proud of you! 💙 #Columbia2025
@bleilab.bsky.social @khanhndinh.bsky.social @elhamazizi.bsky.social

this is probably not the complete picture of KD, but i can definitely sleep better after writing down and confirming this minimal working explanation.

arXiv: arxiv.org/abs/2505.13111

(3/4)

I received a review like this five years ago. It’s probably the right time now to share it with everyone who wrote or got random discouraging reviews from ICML/ACL.

First 11 chapters of RLHF Book have v0 draft done. Should be useful now.

Next:
* Crafting more blog content into future topics,
* DPO+ chapter,
* Meeting with publishers to get wheels turning on physical copies,
* Cleaning & cohesiveness
rlhfbook.com

🔥 Benchmark Alert! MotifBench sets a new standard for evaluating protein design methods in motif scaffolding.
Why does this matter? Reproducibility & fair comparison have been lacking—until now.
Paper: arxiv.org/abs/2502.12479 | Repo: github.com/blt2114/Moti...
A thread ⬇️

The HuggingFace/Nanotron team just shipped an entire pretraining textbook in interactive format. huggingface.co/spaces/nanot...

It’s not just a great pedagogic support, but many unprecedented data and experiments presented for the first time in a systematic way.

I just wanted to see what it looked like 😭

Good God, please. I just want some gradients that don't vanish 😭

I was hoping that recent events would lead to a mass exodus from X. Many have left, but most of the ML and LLM people have not.

I have lost a lot of respect for the ML community.

Now that bluesky has gifs (it didn't work?), I can share (again) my educational notebook on discrete flow matching (by Itai Gat et al.). Also please check the original article and official implementation by Meta!

🐍 github.com/gle-bellier/...
🐍 github.com/facebookrese...
📄 arxiv.org/abs/2407.15595

Really excited about this! We note a connection between diffusion/flow models and neural/latent SDEs. We show how to use this for simulation-free learning of fully flexible SDEs. We refer to this as SDE Matching and show speed improvements of several orders of magnitude.

arxiv.org/abs/2502.02472

I have a sinking feeling that by 2029 I'm going to be faking a British accent so no one will think I was one of the *Americans* working on AI during the regime.

NGL, it's kind of surprising that more people haven't migrated here, especially given what Musk has been doing these days. I don't get it.

Since everyone wants to learn RL for language models now post DeepSeek, reminder that I've been working on this book quietly in the background for months.

Policy gradient chapter is coming together. Plugging away at the book every day now.

rlhfbook.com/c/11-policy-...

Please stop anthropomorphizing language models, it makes them feel really bad

This comments section is the first time I've felt even a shred of hope in eight days.

democracy2025.org/response-center keeps track of it.

Nazi salutes and speaking at neo-Nazi rallies seems bad. There's history that we should learn from.

Something I really like about NLP research is that it makes everything super intuitive. This week I have been thinking about variational inference in NLP and a lot of the things that seemed to require mathematical intuition just become trivial when thinking about language. So cool:)

Tests!! :)

But the memory needed for the value function kills the ones that don't have good GPUs 😭

New randomized, controlled trial by the World Bank of students using GPT-4 as a tutor in Nigeria. Six weeks of after-school AI tutoring = 2 years of typical learning gains, outperforming 80% of other educational interventions.

And it helped all students, especially girls who were initially behind.

I mostly use copilot for writing code (as auto complete), gpt4-o for boiler plate, and o1 for serious debugging or boilerplate with some complexity or a lot of requirements. I also use o1 for quick but slightly involved experiments but not as often.

I use chatgpt over Google for a lot of things because it is really good at fuzzy queries + data aggregation from many sources. I feel that as long as you double check results it is much faster and convenient.