Interesting bit from the llama technical report: not only do data centres for AI training runs draw huge loads on the grid, the loads can be uncharacteristically spiky.

My overpriced health tracker really does it all: helping me on my physical performance *and* my code’s performance 💪

new from me: I’ve teamed up with Anastasia Bektimirova to explore why so much policy work is basically pointless. We walk through the mistakes tech companies, consultancies, think tanks, and lobbyists make, how it impoverishes debate, and how we can improve it. 🧵

Fantastic talk from Richard Sutton on his work on plasticity loss in deep continual learning (nature.com/articles/s4158…) at ICARL this afternoon. Will add video link in replies

How do transformers learn arithmetic tasks, such as GCD and modular sums and products? My talk in Collège de France on November 4th (in French, but the English subtitles are quite good). Thank you Timothy Gowers @wtgowers for inviting me to your seminar! youtube.com/watch?v=e0jUi8…

Funny bug in our coding agent: the time in the prompt was 6pm on a Friday, and given some persistent issues, the agent decided to clock out for the weekend.

Ever wondered why presenting more facts can sometimes *worsen* disagreements, even among rational people? 🤔 It turns out, Bayesian reasoning has some surprising answers - no cognitive biases needed! Let's explore this fascinating paradox quickly ☺️

You either out-innovate the competition in the race to AGI, or you get Schmidhubered trying

le contrat artificiel

arXiv -> TeX source -> paste into o3 -> turn on voice mode -> chat about the paper with a middling PhD student who can search stuff on the web in real-time

I'd say I'm both an ML true believer and an AI skeptic: super hyped about the concept of extrapolation from compressed distributed representation of statistical patterns in data, think it's useful and philosophically significant, not sure it's good at building genies

I love it when authors explain (their own) prior papers in a related work section. It's often the crispest, no-bs, 2-sentence TLDR pointing out the key method and intuition

If you give your AI model a French name, it is perhaps not surprising it will be offline 20% of the year.

🚨New Paper!🚨 We trained reasoning LLMs to reason about what they don't know. o1-style reasoning training improves accuracy but produces overconfident models that hallucinate more. Meet RLCR: a simple RL method that trains LLMs to reason and reflect on their uncertainty --