Parallelizing "inherently sequential" processes has become all the rage in the era of GPUs. But can we really parallelize anything? In work led by me and @Leokoz8, we show that the "predictability" of a dynamical system determines whether it can be parallelized efficiently.

When we began applying diffusion to language in my lab at Stanford, many doubted it could work. That research became Mercury diffusion LLM: 10X faster, more efficient, and now the foundation of @_inception_ai. Proud to raise $50M with support from top investors.

Amazing work from a great team

Folks say RNNs can’t be parallelized, but they can! This cool new work shows which nonlinear systems can be parallelized, and that you can parallelize MCMC **within a chain** !

Love this line of work from @scott_linderman @xavierjgonzalez @dmzoltowski and others on parallelizing nonlinear systems. The new papers explore parallelizing MCMC as well as characterizing the types of nonlinear systems that can or cannot be parallelized.

Excited to share these two papers on parallelizing MCMC and theory for parallelizing nonlinear sequence models! This is a fun line of work and there is lots more to explore in these areas. Collabs with @SkylerWu9 @xavierjgonzalez @Leokoz8 Ken Clarkson & @scott_linderman

We've raised $100M from Kleiner Perkins, Index Ventures, Lightspeed, and NVIDIA. Today we're introducing Sonic-3 - the state-of-the-art model for realtime conversation. What makes Sonic-3 great: - Breakthrough naturalness - laughter and full emotional range - Lightning fast -

As part of our recent work on memory layer architectures, I wrote up some of my thoughts on the continual learning problem broadly: Blog post: https://t.co/HNLqfNsQfN Some of the exposition goes beyond mem layers, so I thought it'd be useful to highlight separately:

I think we're going to start seeing more continual learning techniques like this one that dynamically adapt which parameters are updated -- some really important ideas in here

Want the freedom of a fancy fellowship, but not the year-long wait or arduous application? Come join my lab! Work on neuroscience and AI, explore your creativity, be independent or work closely with me, collaborate widely, and have a lot of fun! https://t.co/zmROCA5Ib6

We at TMLR are proud to announce that selected papers will now be eligible for an opportunity to present at the joint NeurIPS/ICML/ICLR Journal-to-Conference (J2C) Track: https://t.co/CyjZtqbnBS

Today, we’re launching Parsed. We are incredibly lucky to live in a world where we stand on the shoulders of giants, first in science and now in AI. Our heroes have gotten us to this point, where we have brilliant general intelligence in our pocket. But this is a local minima. We

🚨New paper! Generative models are often “miscalibrated”. We calibrate diffusion models, LLMs, and more to meet desired distributional properties. E.g. we finetune protein models to better match the diversity of natural proteins. https://t.co/2c06vD0x2D https://t.co/9Tbhf6ml8K

My thesis, 𝘈 𝘵𝘩𝘦𝘰𝘳𝘺 𝘰𝘧 𝘵𝘩𝘦 𝘤𝘰𝘮𝘱𝘶𝘵𝘢𝘵𝘪𝘰𝘯𝘢𝘭 𝘱𝘰𝘸𝘦𝘳 𝘢𝘯𝘥 𝘭𝘪𝘮𝘪𝘵𝘢𝘵𝘪𝘰𝘯𝘴 𝘰𝘧 𝘭𝘢𝘯𝘨𝘶𝘢𝘨𝘦 𝘮𝘰𝘥𝘦𝘭𝘪𝘯𝘨 𝘢𝘳𝘤𝘩𝘪𝘵𝘦𝘤𝘵𝘶𝘳𝘦𝘴, is now online:

Bit late now, but happy to announce that "parallelizable Linear Source Transition Mark networks" #pLSTM was accepted at the main track of NeurIPS 2025! If you cannot make it to San Diego, watch the talk at the ASAP Seminar: https://t.co/RmUCYhos5E

Really enjoyed presenting our @NeurIPSConf papers to the asap seminar! With @LeoKoz8. Here is the talk on youtube:

Super excited to be presenting my latest @NeurIPSConf paper "Predictability enables Parallelization of nonlinear SSMs" with @Leokoz8 at the ASAP seminar at 2 pm eastern tomorrow Tuesday 9/30. Tune in if you are interested! zoom link: https://t.co/sYyzzMOvoL asap seminar:

@X I think @X requires that videos (like .mp4's) be vertical? The error messages could be a lot more helpful about explaining what aspect ratio is required though.