🎓 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.

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:)

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.

Does anyone have any good resources to learn about quantization? Any essential papers to read and resources about how to use/quantize models in practice are greatly appreciated!

1-> 2 -> 3 -> 3.5 -> 4 -> 4o -> o1 -> o3 I guess we need AGI just to figure out how to name things

If you are into ML theory (RL or not) with a proven track record, and you are interested in an industry research position, PM me. Feel free to spread the word.

🧵 Excited to share #Echidna, a Bayesian framework for quantifying the impact of gene dosage on phenotypic plasticity: tinyurl.com/296kf7hf! With @elhamazizi.bsky.social and @mingxz.bsky.social, we integrate scRNA-seq & WGS to uncover how CNAs drive tumor evolution and transcriptional variability.

Proud of this work spearheaded by the phenomenal @jlfan.bsky.social and @mingxz.bsky.social in collaboration w/ Ben Izar! The past 3 years we've worked hard to unravel how #CNVs shape #tumor phenotypic plasticity seen in #singlecell #RNAseq data ➡️ #Echidna 🦔

Hello! We will be presenting Estimating the Hallucination Rate of Generative AI at NeurIPS. Come if you'd like to chat about epistemic uncertainty for In-Context Learning, or uncertainty more generally. :) Location: East Exhibit Hall A-C #2703 Time: Friday @ 4:30 Paper: arxiv.org/abs/2406.07457

I'm on my way to #NeurIPS2024. On Friday I'm going to present my latest paper with Yuval Benjamini. The gist is in the comments, and come chat with me to hear more!

The circuit hypothesis proposes that LLM capabilities emerge from small subnetworks within the model. But how can we actually test this? 🤔 joint work with @velezbeltran.bsky.social @maggiemakar.bsky.social @anndvision.bsky.social @bleilab.bsky.social Adria @far.ai Achille and Caro

I have been a little bit scarce on social media in the last few months. Some of that has just been from being busy at work, but some of it has had to do with my father's passing. He hated the very idea of social media, but he religiously followed my Twitter, and then Bluesky, feeds. /n

We are thrilled to share #Decipher 🔍! Extremely proud of Achille and Joy for leading the development of this creative #ML tool combining VAEs with deep exponential models for comparing disease & healthy trajectories ➕ applying it to study leukemic onset! biorxiv.org/conte…

I wanted to make my first post about a project close to my heart. Linear algebra is an underappreciated foundation for machine learning. Our new framework CoLA (Compositional Linear Algebra) exploits algebraic structure arising from modelling assumptions for significant computational savings! 1/4

Thank you Nature and @anilananth.bsky.social for this great feature on LLMs and AGI (and for highlighting our work arxiv.org/abs/2406.03689)

From our lab finally I convinced Yuli Slavutsky (generalization/uncertainty/robustness) @yulislavutsky.bsky.social and Claudia Shi (science of LLMs and AI safety) @claudiashi.bsky.social to use bluesky. Go give them a follow! 😁

Very happy to share some recent work by my colleagues @velezbeltran.bsky.social, @aagrande.bsky.social and @anazaret.bsky.social! Check out their work on tree-based diffusion models (especially the website—it’s quite superb 😊)!

I am very excited to share our new Neurips 2024 paper + package, Treeffuser! 🌳 We combine gradient-boosted trees with diffusion models for fast, flexible probabilistic predictions and well-calibrated uncertainty. paper: arxiv.org/abs/2406.07658 repo: github.com/blei-lab/tre... 🧵(1/8)

A common question nowadays: Which is better, diffusion or flow matching? 🤔 Our answer: They’re two sides of the same coin. We wrote a blog post to show how diffusion models and Gaussian flow matching are equivalent. That’s great: It means you can use them interchangeably.

How to drive your research forward? “I tested the idea we discussed last time. Here are some results. It does not work. (… awkward silence)” Such conversations happen so many times when meetings with students. How do we move forward? You need …

I wrote a summary of the main ingredients of the neat proof by Hugo Lavenant that diffusion models do not generally define optimal transport. github.com/mathematical...

Look if ClosedForProfitAI wants to train ChatPilotLLM on bluesky, they're not gonna use 1 or 2 million posts from huggingface they'll just download the entire firehose, have extremely skilled people update the model as new posts get downloaded, and not tell anybody what data they're using to train.

It's pretty sad to see the negative sentiment towards Hugging Face on this platform due to a dataset put by one of the employees. I want to write a small piece. 🧵 Hugging Face empowers everyone to use AI to create value and is against monopolization of AI it's a hosting platform above all.

FYI, here's the entire code to create a dataset of every single bsky message in real time: ``` from atproto import * def f(m): print(m.header, parse_subscribe_repos_message()) FirehoseSubscribeReposClient().start(f) ```

Rather than restricting data to only the richest and most powerful (as reddit, facebook, and twitter do), bsky makes it available to everyone. Personally, I think that's a good thing.

I've never understood someone less than the journalists who snarkily write about how no one could conceivably ever want any AI tool. I can't steelman their position. I couldn't pass an ideological turing test. I cannot put myself in their shoes. It's very frustrating.

I'm disheartened by how toxic and violent some responses were here. There was a mistake, a quick follow up to mitigate and an apology. I worked with Daniel for years and is one of the persons most preoccupied with ethical implications of AI. Some replies are Reddit-toxic level. We need empathy.

Excited to share OLMo 2! 🐟 7B and 13B weights, trained up to 4-5T tokens, fully open data, code, etc 🐠 better architecture and recipe for training stability 🐡 staged training, with new data mix Dolmino🍕 added during annealing 🦈 state-of-the-art OLMo 2 Instruct models #nlp #mlsky links below👇

"i can fix her 🥹" meme but it's me and tokenizers

On the TCS job market: Juspreet Singh Sandhu! Optimization, complexity & probabilistic analysis of random CSPs, and models in classical & quantum statistical physics. Techniques: random matrices, free probability, concentration inequalities, stochastic analysis. 1/2 #TCSSky #AcademicJobMarket