I'm starting a machine learning society at my uni :) We've partnered with the University of Gloucestershire to get members access to the on-site B200 cluster, to start the year we are going to recreate GPT-2 I'm very excited! hope to be able to share more soon

I am going to be in SF till the 13th, if you are interested in talking to me please reach out, my goal for this trip is to meet as many people in ml research and engineering as possible. I have just updated my personal website (noah dot jp dot net) with the relevant information!

I did it. I reached Highway 101 leading into downtown San Francisco with all the AI SaaS ads. This must be what Hannibal felt like after he cleared the Alpine Pass

🔍 New paper: How do vision-language models actually align visual- and language representations? We used sparse autoencoders to peek inside VLMs and found something surprising about when and where cross-modal alignment happens! Presented at XAI4CV Workshop @ CVPR 🧵 (1/6)

Sparse autoencoders (SAEs) can be used to elicit strong reasoning abilities with remarkable efficiency. Using only 1 hour of training at $2 cost without any reasoning traces, we find a way to train 1.5B models via SAEs to score 43.33% Pass@1 on AIME24 and 90% Pass@1 on AMC23.

In particular, I think the current wave of SAE skepticism is about as irrational as the wave of SAE hype that preceded it. It is wholly plausible that there are a set of improvements that will lead to massive improvements. Just, those improvements won't be found by throwing

🚨 New paper alert! Linear representation hypothesis (LRH) argues concepts are encoded as **sparse sum of orthogonal directions**, motivating interpretability tools like SAEs. But what if some concepts don’t fit that mold? Would SAEs capture them? 🤔 1/11

How does this affect LeBron legacy?

hey its me :)

https://t.co/jYtSjS2z7m

am i sure the death star is going down? look at my quant. look at him! you notice anything different about him? look at his eyes. i’ll give you a hint—his name’s a fucking number!! he doesn’t even speak english—it’s all beep-boop shit!! yeah, i’m sure.

We created a canvas that plugs into an image model’s brain. You can use it to generate images in real-time by painting with the latent concepts the model has learned. Try out Paint with Ember for yourself 👇

Dont forget learning dynamics of fine tuning and what SFT does to latent space vs DPO either, i think something is brewing…

A lot of alpha in this graph remember “Large Language Models are Zero-Shot Reasoners” as well? I love it when a plan comes together

🚀 Excited to share the most inspiring work I’ve been part of this year: "Learning to Reason without External Rewards" TL;DR: We show that LLMs can learn complex reasoning without access to ground-truth answers, simply by optimizing their own internal sense of confidence. 1/n

The more I look into the system card, the more I see over and over 'oh Anthropic is actually noticing things and telling us where everyone else wouldn't even know this was happening or if they did they wouldn't tell us.'

Even when tasks feel bottlenecked on people (e.g. in lab throughput), it’s not a manpower problem that robotics will fix. Lab automation is expensive, limited in scope, and brittle. You’re not going to run 10k lab experiments overnight, no matter how many robot arms you buy.

🧵For Claude Opus 4, we ran our first pre-launch model welfare assessment. To be clear, we don’t know if Claude has welfare. Or what welfare even is, exactly? 🫠 But, we think this could be important, so we gave it a go. And things got pretty wild…

I’m stoked to share our new paper: “Harnessing the Universal Geometry of Embeddings” with @jxmnop, Collin Zhang, and @shmatikov. We present the first method to translate text embeddings across different spaces without any paired data or encoders. Here's why we're excited: 🧵👇🏾

Introducing Multiverse: the first AI-generated multiplayer game. Multiplayer was the missing piece in AI-generated worlds — now it’s here. Players can interact and shape a shared AI-simulated world, in real-time. Training and research cost < $1.5K. Run it on your own PC. We