Anurag @buildingcells

how much can models remember? there’s a neat perspective on neural networks that deserves to be more widely known here’s how you could have invented every layer of the transformer from memory (1/n) from my new blog post!

this is so so somatic wow wow wow the models have space to discretize more concepts than our languages developed words for, it’s like the ultimate agglutinative language

Nick @nickcammarata

a week ago

interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came

878 3K 20K 4.5M 20K

this article is just so well written, it's altering my brain chemistry

Nick @nickcammarata

a week ago

interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came

878 3K 20K 4.5M 20K

interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works already-clean.goodfire.com

Can’t wait to dig in tonight, will I really feel the “move 37”

Octi Zhang @octi_zhang

2 weeks ago

(1/6) From @GoogleDeepMind’s AlphaGo to AlphaZero, we saw machines find moves that reached beyond what we’d imagined for them. No one taught them those strategies. They emerged from reinforcement learning: playing, failing, and trying again under a single binary task reward for

64 335 3K 415K 2K

@aliansarinik Sent an email along

This is dope

Georgia Channing @cgeorgiaw

2 weeks ago

Today, we're releasing the Carbon Annotation Database and the model behind it, Carbon-A. We've used this model to discover 566.34 million new gene candidates across 22,617 species, thousands of which have never been studied like this before. We wet-lab validated a number of

12 59 329 41K 159

Feeling FOMO, working on advancing real world applications of scientific progress and improve lives in the most tangible way is the best work there is!

Molly O’Shea @MollySOShea

3 weeks ago

BREAKING: The U.S. government, Google, Meta & Zuckerberg-backed @biohub are launching a $1.8B push to build the data layer for AI biology. The goal is to create the largest open, AI-ready biological dataset ever assembled, giving models the data needed to better predict how

keyboard_arrow_left Previous keyboard_arrow_right Next

144 523 4K 1.8M 1K

This is a huge step in the right direction

Jiaheng Hu @JiahengHu1

3 weeks ago

We’ve all seen the cool “Astra for robotics” demos. But behind the scenes? Slow rollouts, hand-engineered APIs, and/or simulated behaviors that struggle on real robots. How do we make coding agent actually work in the physical world? 🦾 Introducing SimEX: Simulation-Integrated

11 60 296 32K 212

Exactly why using RLCD models such as Jev, etc to help decide when to have operators intervene could be incredibly valuable! x.com/buildingcells/…

Junfan Zhu 朱俊帆 ✈️ CoRL @junfanzhu98

3 weeks ago

The winning robot stack will not be the one that never needs a human. It’ll be the one where a human in NY can grab a failure in SF, finish the grasp, hand control back—and that recovery can become the next training example. @cortexairobot Harness is that layer. Prompt. Steer.

0 4 14 984 2

More pre-training doesn't necessarily mean better generalization? This new paper discovers mode-hopping, where LLMs repeatedly switch between shallow pattern-matching and actual generalization, even while training loss remains stable. For example, OLMo3-32B went from 81% accuracy to 0%, then back to 81.7% within just 40B training tokens. On top of that, selecting an earlier 4.5T-token checkpoint instead of a 4.9T one improved GPQA transfer after math fine-tuning (36.3% vs 29.8%) and robustness to alignment attacks (53% vs 21%) alphaxiv.org/abs/2609.33150

Pulled almost an all-nighter, but I finally got done writing my blog on reinforcement learning. Next part of the CUDA blog next Friday!!... hopefully... no promises :X x.com/i/article/2106…

How a robot arm is controlled, explained for ML people new to robot learning. First of the explainers from my own speedrun. Next one will be on ACT. Follow me to catch it. x.com/i/article/2100…

Instead of training to predict the next action, ACT (@tonyzzhao et al.) trains to generate the next 100. How and why, explained for ML people new to robot learning. Second from my speedrun. Next one, on Diffusion Policy, next week. Follow me to catch it. x.com/i/article/2102…

How diffusion and flow matching work, built up from the VAE, as used by Diffusion Policy (@chichengcc et al.), all GR00T models, and most π models. For ML people new to robot learning. x.com/i/article/2105…

Planning over long horizons is difficult when every action must be predicted one step at a time. FlexiWorld addresses this by learning flexible action chunks—groups of primitive actions that can vary in length—while training on goals spanning multiple time scales. The JEPA-based world model combines mixed-span goal supervision, a variable-length causal action encoder, an autoregressive actor, and Student Forcing to reduce errors from generated action prefixes. Its ARCEM planner searches over action adjustments while using feedback within each chunk and latent predictions at chunk boundaries. Across four benchmarks and goal distances, FlexiWorld with ARCEM reaches 89.29% mean success, compared with 83.98% for the strongest baseline. It also supports different planning chunk lengths without retraining; longer chunks speed up planning by about 1.3× on average while maintaining comparable success. Title: FlexiWorld: Learning and Planning via Flexible Action Chunks Across Multiple Time Scales Project URL: shidu-ren.github.io/FlexiWorld-Pro… Paper: arxiv.org/abs/2609.35138…