Simons Institute for the Theory of Computing

2/2 "Most of the interesting computations that happen today are in some way or another computations that happen on individuals’ potentially sensitive data," said Katrina Ligett of HUJI at the Simons Institute workshop on Theory of Computing and Healthcare. simons.berkeley.edu/talks/katrin...

1/2 "There is sort of an obligation on the part of somebody who spends a lot of time thinking about privacy to open with the bad news," said Katrina Ligett of HUJI, talking about Research on Sensitive Data at the Simons Institute workshop on Theory of Computing and Healthcare.

Join us at 3:30 p.m. PT. Register to attend or access the livestream.

simons.berkeley.edu/events/respo...

3/3 "Current training recipe doesn't really support fragmented data," said Sewon Min of @ucberkeleyofficial.bsky.social at the Simons Institute workshop on Federated and Collaborative Learning Boot Camp. Video: youtu.be/vmt2_LZ8zgI?...

2/3 One reason neural scaling laws might stall is because data is getting fragmented. "I’m talking about proprietary datasets, where datasets are owned by different owners...that cannot be gathered into a central location," said Sewon Min of @ucberkeleyofficial.bsky.social at the Simons Institute.

1/3 Will increasing compute and training data continue to improve performance of foundation models? Maybe. "I’m interested in asking: if it’s not the case, then why?" said Sewon Min of @ucberkeleyofficial.bsky.social at the Simons Institute workshop on Federated and Collaborative Learning Boot Camp

YouTube video by SparX by Mukesh Bansal Why Machines Learn: The Elegant Math Behind AI with Anil Ananthaswamy | SparX by Mukesh Bansal

#FlashbackFriday

www.youtube.com/watch?v=soQR...

UC Berkeley and Project CETI study shows sperm whales communicate in ways similar to humans The way sperm whales communicate may be more similar to human language than previously thought. The acoustic properties of whale calls resemble vowels, a defining feature of human language, according ...

#FlashbackFriday

ls.berkeley.edu/news/uc-berk...

Could AI Supplant a Mathematician? Professor Andrew Blumberg is part of a team that tested AI’s upper limits by asking it to answer unsolved math problems.

#1stproof

news.columbia.edu/news/could-a...

YouTube video by Simons Institute for the Theory of Computing Alignment Problems in AI GovernanceLocation

Next was a great talk by @r-jy.bsky.social and Greg Demirchyan on alignment problems in AI governance at @simonsinstitute.bsky.social www.youtube.com/watch?v=rV2P... (5/7)

Congratulations to our colleague John Wright, who has received a 2026 Sloan Fellowship!

chemistry.berkeley.edu/news/seven-u...

4/4 Trained neural networks can also leak private information, for example, via "fill-in-the-blank" and secret-sharer attacks, said said Om Thakkar of OpenAI, at the Simons Institute workshop on Federated and Collaborative Learning Boot Camp. Video: simons.berkeley.edu/talks/om-tha...

3/4 "Leakage can come from [ML] algorithms themselves...For example, gradients can leak a lot of information," said Om Thakkar of OpenAI, at the Simons Institute workshop on Federated and Collaborative Learning Boot Camp. Video: simons.berkeley.edu/talks/om-tha...

2/4 "Training data can get leaked due to unauthorized access," because of internal adversaries or external attackers, said Om Thakkar of OpenAI, at the Simons Institute workshop on Federated and Collaborative Learning Boot Camp. Video: simons.berkeley.edu/talks/om-tha...

1/4 Where can data leakage happen in a machine learning pipeline? "Leaks can happen at any point in the pipeline, said Om Thakkar of OpenAI, at the Simons Institute workshop on Federated and Collaborative Learning Boot Camp. Video: simons.berkeley.edu/talks/om-tha...

Testing Artificial Mathematical Intelligence YouTube video by Simons Institute for the Theory of Computing

Inspired by Turing's "Computing machinery and intelligence," Emily Riehl proposes a series of tests to help identify whether a generative AI system can meaningfully contribute to the process of doing mathematics.

youtube.com/live/svF-1ek...

Join us February 24 for the second and final Richard M. Karp Distinguished Lecture of the semester. Registration is required.

simons.berkeley.edu/events/respo...

3/3 "It’s estimated that there’s 2,000 trillion tokens that are private today": Patrick Foley of Flower Labs at the Simons Institute workshop on Federated and Collaborative Learning Boot Camp. FL could enable secure access to this data to train future LLMs. simons.berkeley.edu/talks/cheste...

2/3 "Yes, absolutely seeing [federated learning] having a role 5 yrs from now, based on the amount of data that is estimated to be private versus what is known to be public," said Patrick Foley of Flower Labs. Today's best LLMs have been trained on 5-15 trillion tokens of data.

1/3 If we get to AGI in 5 years, will we need federated learning, i.e. decentralized ML for training models over multiple nodes that each have private data? Q posed at the Simons Institute workshop on Federated and Collaborative Learning Boot Camp. Video simons.berkeley.edu/talks/cheste...

Mathematicians issue a major challenge to AI—show us your work Frustrated by the AI industry’s claims of proving math results without offering transparency, a team of leading academics has proposed a better way

More on #1stproof, from @sciam.bsky.social

www.scientificamerican.com/article/math....

Distinct AI Models Seem To Converge On How They Encode Reality | Quanta Magazine Is the inside of a vision model at all like a language model? Researchers argue that as the models grow more powerful, they may be converging toward a singular “Platonic” way to represent the world.

If you've been feeling that what's missing from the discussion of AI is a bit of Plato, check out @benbenbrubaker.bsky.social on the @quantamagazine.bsky.social podcast (from last month).

www.quantamagazine.org/distinct-ai-...

These Mathematicians Are Putting A.I. to the Test

From @nytimespr.bsky.social, a closer look at #1stproof, a community experiment to see how well AI can do research math.

www.nytimes.com/2026/02/07/s...

When you do the math, humans still rule — Harvard Gazette Harvard’s Lauren Williams, a MacArthur “genius,” joins international effort to challenge notions of AI supremacy.

news.harvard.edu/gazette/story/2026/02/when-you-do-the-math-humans-still-rule/

#1stproof

Theory at the Institute and Beyond, February 2026 Let me tell you about two lists of questions — one for humans, and one for AI — and how they were made.

In the latest installment of Theory at the Institute and Beyond, Senior Scientist Nikhil Srivastava explores innovative approaches to workshop design, and introduces a community experiment to see how well AI can do research math.

simons.berkeley.edu/news/theory-...

#1stproof

Theory of Computing and Healthcare The Theory of Computing and Healthcare Workshop at Berkeley aims to bridge innovative concepts from theoretical computer science and the pressing, real-world challenges of modern healthcare. As health...

Next week at the Simons Institute, a workshop on Theory of Computing and Healthcare. Join us!

simons.berkeley.edu/workshops/th...

Join us Thursday!

simons.berkeley.edu/events/align...

Just how big are today's LLM pre-training data centers? It'd have taken 5.5 years to train Grok3 on El Capitan, the world's fastest supercomputer, said Keith Rush of Google DeepMind at the Simons Institute workshop on Federated and Collaborative Learning Boot Camp simons.berkeley.edu/talks/keith-...

Alignment Problems in AI Governance In an event comprising short talks and dialogue, Simons Institute Law and Society Fellows Rui-Jie Yew and Greg Demirchyan will explore two challenges of alignment in AI governance. First, we currently...

Thurs., 2/5 at 3:30 p.m., Law & Society Fellows Rui-Jie Yew and Greg Demirchyan discuss Alignment Problems in AI Governance.

simons.berkeley.edu/events/align...

Join us today at 3:30. Registration is required.

simons.berkeley.edu/events/lets-...