ReLU is fondly named after Relu Patrascu who put together GPU clusters in the early 2010s accelerating neural net research at U of T Department of Computer Science

will math journals move towards requiring authors to submit formalized computer proofs? newscientist.com/article/246189…

americans’ satisfaction with democracy: on life support

An interesting aspect of this is, with all the recent activity around LLM reasoning in formal proof systems (“symbol pushing”), people seem to miss that many mathematicians think through a problem by visualizing equations and abstract objects.

‘We are now confident we know how to build AGI as we have traditionally understood it.’ This is in no small part due to OpenAI’s early research investments in reinforcement learning. Mr. Altman’s recent comments on not knowing what side of the singularity we are in, I think,

Chinese people are smart, not just copying others. Expect more Deepseek like advances in AI scholarship and engineering.

David Hilbert envisioned a “complete” mathematics, a firm foundation from which all mathematical truths could be derived. But researchers later showed that math is actually riddled with undecidable statements. Another important problem has now been deemed undecidable, according

👍

. NVIDIA’s Jensen Huang shares how AlexNet — developed by U of T Department of Computer Science alumni Alex Krizhevsky and Ilya Sutskever and @uoft University Professor Emeritus Geoffrey Hinton — inspired the company to go “all in” on autonomous vehicles. ⤵️

Three papers with authors from @uoftcompsci are among the 25 most cited of the 21st century, according to a recent analysis by @nature 📄 uoft.me/bwH

The brain decides if a movement belongs to you before you even make it. Fascinating Nature Communications study reveals that pre-movement alpha rhythms in motor cortex are not just preparatory, they’re the foundation of subjective control and sense of agency.

How harmful is GenAI around elections? Will it trigger a misinformation apocalypse and upend elections? I am happy to finally be able to share Sacha Altay’s & my answers to these and other questions on which we have been working for a year and which is out via Knight First Amendment Institute.

Nice result but what does it mean for "everyday" reasoning abilities? DM's system was over-engineered to work well on the IMO domain, having access to past problems & solutions, hints & tips, and a sizable dev team tweaking every system detail. It's hard to imagine such a

short paper by Yoshua Bengio and me about what LLMs tell us about our own mental lives (extended from earlier blog post) philpapers.org/rec/YOSWDL

It appears there’s substantial ‘overfitting’ to existing reasoning benchmarks such that they fail badly on never before seen ones.

ELI5⃣ 👉 When your LLM inference job is mixed with a batch of other peoples requests, GPU kernels doing reduction, matmul, attention can be affected by the batchsize, and that causes small numerical differences. 🔥Why is it important? ... 👏Horace He and team proposes a set

What a great department!

I recommend checking out Chang Xu happy hour if you’re attending Neurips this year!