It's worth noting that a DeepThink system with no access to this corpus also got gold (again according to the official graders), with exactly the same score.
How does this simulation work, since floating point numbers are all rational? Is there some notion of effective irrationality where a floating-point number can't be expressed as a ratio of integers within the floating point system?
Today we're putting out an update to the JAX TPU book, this time on GPUs. How do GPUs work, especially compared to TPUs? How are they networked? And how does this affect LLM training? 1/n
New work out on arXiv (arxiv.org/abs/2010.15114) about the geometry of integration in RNNs. This was led by Vinay Ramasesh (@vinayramasesh) and Kyle Aitken (@kyle__aitken), with additional co-authors Ankush Garg, Yuan Cao, and David Sussillo (@sussillo). #tweeprint below! 🧵
Just a note, many of the researchers involved with this work were also heavily involved with the BIG Bench effort which hopes to alleviate some of these issues: github.com/google/BIG-ben…
Multiple AI systems won gold medals at the 2025 International Mathematical Olympiad (IMO). Exciting as that sounds, @GregHBurnham argues that it represents little progress: an unlucky draw of problems made the event relatively uninformative.
Is that cope? Judge for yourself. 🧵
Another resource that might be helpful towards teaching this line of work is our recent @DepthFirstLearn course covering, among other topics, this mean-field treatment of wide neural networks: depthfirstlearning.com/2020/Resurrect…
Very interesting! Question: from the article it sounds like semaglutide has two mechanisms: 1) lowering cravings for highly-processed food, and 2) lowering the bodyfat setpoint. Are these independent? Or is (1) a consequence of (2)?
