The world's leading venue for collaborative research in theoretical computer science. Follow us at http://YouTube.com/SimonsInstitute.
@simonsinstitute.bsky.social
Simons Institute for the Theory of Computing
@simonsinstitute.bsky.social
@simonsinstitute.bsky.social@simonsinstitute.bsky.socialEarlier this month, we hosted a working group to think about how the TCS community should respond to AI progress. Their report, released on Monday, highlights 12 near-term actions that received broad support among the participants.
@ccanonne.github.ioReposted by @simonsinstitute.bsky.social"Well, that's just, like, your opinion, man."
(remotely attending the Simons Institute's Karp Distinguished Lecture by @booleananalysis.bsky.social) @simonsinstitute.bsky.social
@simonsinstitute.bsky.socialEarlier this month, we hosted a working group to think about how the TCS community should respond to AI progress. Their report, released yesterday, highlights 12 near-term actions that received broad support among the participants.
@simonsinstitute.bsky.social4/4 "This allows you to interpolate between auto-regression and diffusion," said Volodymyr Kuleshov of @cornelluniversity.bsky.social at the Simons Institute workshop on Diffusion Generative Modeling: Progress and Next Steps. Video: simons.berkeley.edu/talks/volody...
@simonsinstitute.bsky.social3/4 Block diffusion addresses this limitation. "You'd run diffusion over blocks of tokens in parallel and generate the following block of tokens, using a fixed length diffusion block, conditioned on the [generated] tokens," said V. Kuleshov of @cornelluniversity.bsky.social at the Simons Institute
@simonsinstitute.bsky.social2/4 For e.g., current diffusion language models have "been defined to only generate fixed length sequences," whereas autoregressive models can keep adding tokens and generate sequences of variable length, said Volodymyr Kuleshov of @cornelluniversity.bsky.social at the Simons Institute
@simonsinstitute.bsky.social1/4 Are diffusion language models ready for the real world? Not quite. Today's diffusion language models are missing some key ingredients, said Volodymyr Kuleshov of @cornelluniversity.bsky.social, at the Simons Institute workshop on Diffusion Generative Modeling: Progress and Next Steps
@ccanonne.github.ioReposted by @simonsinstitute.bsky.socialQuite the exciting talk by Ryan O'Donnell (@booleananalysis.bsky.social) at the Simons Institute next week, on Expander Graphs.
We'll finally know which ones are his favorites!
Livestream option available upon (free) registration: simons.berkeley.edu/events/my-fa... @simonsinstitute.bsky.social
@simonsinstitute.bsky.socialJoin us Tuesday, 9/29 for the first Richard M. Karp Distinguished Lecture of this academic year, featuring @booleananalysis.bsky.social.
@simonsinstitute.bsky.socialWe invite your project proposals for Circles, the Simons Institute – Jane Street Small Group Collaborations, which supports groups of 3-6 researchers for 4 weeklong gatherings over 2 years.
Apply by Oct. 15 (deadline extended).
@simonsinstitute.bsky.social5/5 Volodymyr Kuleshov of @cornelluniversity.bsky.social spoke at the Simons Institute workshop on Diffusion Generative Modeling: Progress and Next Steps. Video: simons.berkeley.edu/talks/volody....
@simonsinstitute.bsky.social4/5 "To carry [over the pre-training gains] to post-training and inference time scaling...we need to develop language models that are fully parallel both in training and inference," said Volodymyr Kuleshov of @cornelluniversity.bsky.social at the Simons Institute. Diffusion LMs are an option.
@simonsinstitute.bsky.social3/5 "Both of these techniques are heavily bottlenecked by our ability to generate from the model quickly, both in RL and inference time scaling...the current algorithm used for generation is inherently sequential," said Volodymyr Kuleshov of @cornelluniversity.bsky.social at the Simons Institute
@simonsinstitute.bsky.social2/5 But, "most of the gains [in language modeling performance in the last two years] have come from scaling post-training, especially using reinforcement learning as well as inference time scaling," said Volodymyr Kuleshov of @cornelluniversity.bsky.social at the Simons Institute
@simonsinstitute.bsky.social1/5 The case for diffusion language models: "A lot of the [early] gains in language modeling performance have come from scaling pre-training...[the training algorithm] was designed to be very parallelizable across GPUs": Volodymyr Kuleshov of @cornelluniversity.bsky.social at the Simons Institute
@simonsinstitute.bsky.social"Of course, we also have failures," said Michal Irani (@weizmann.ac.il), at the Simons Institute workshop on Topics in Intelligence: World Models and Social Reasoning. For e.g., the fMRI scan of a person viewing a cat became a bear. Video: simons.berkeley.edu/talks/michal...
arxiv.org/abs/2510.25976
@simonsinstitute.bsky.social4/5 Here are some examples of successful reconstructions of the images observed by a person, using only the fMRI images. For each pair, left is the original image, right is the reconstruction from the fMRI scan. Paper: arxiv.org/abs/2510.25976 Roman Beliy et al.
@simonsinstitute.bsky.social3/5 A diffusion model uses both the initial guess & the CLIP embedding to create high-level images that are semantically meaningful and consistent with the global layout of the original image, said Michal Irani (@weizmann.ac.il, at the Simons Institute. Paper: arxiv.org/abs/2510.25976
@simonsinstitute.bsky.social2/5 Given an fMRI image, a high-level semantic decoder learns to predict the CLIP embedding of the image the person saw. A low-level structural decoder learns to predict crude images from the fMRI image, forming an initial guess of the image's global layout, said Michal Irani, @weizmann.ac.il.