Man, I love a good disambiguation. If everyone was familiar with this breakdown (by Don Moore) it would make conversations about overconfidence so much less frustrating. From onlinelibrary.wiley.com/doi/full/10.11…

I made a comment buried in a reddit thread yesterday about how my views on privacy have changed over the last few years (I'm more strongly pro-privacy now), and was encouraged to signal-boost it. Here you go: reddit.com/r/ethereum/com…

A dead fish can swim upstream. Fluid dynamics is metal. x.com/chubicki/statu…

Today in Nature Communications, we explore how AI agents can better communicate and cooperate in Diplomacy - a 7-player board game of coordination and alliance formation. 🤝 Find out more: dpmd.ai/diplomacy-natu…

Using negotiation algorithms, agents can form contracts regarding their next move & outperform others without this ability. When agents may deviate from past contracts, we show that sanctioning peers who break promises fosters more truthful communication: dpmd.ai/diplomacy-tw

Work by: János Kramár, Ian Gemp, Andrea Tacchetti, Kevin McKee, Mateusz Malinowski, Thore Graepel and Yoram Bachrach.

Our latest paper (arxiv.org/abs/2309.02390) provides a general theory explaining when and why grokking (aka delayed generalisation) occurs – a theory so precise that we can predict hyperparameters that lead to partial grokking, and design interventions that reverse grokking! 🧵👇

Recent work on Zoneout, a successful form of dropout regularization for RNNs: arxiv.org/abs/1606.01305

Reverse Twitter: a social network that encourages meaningful debate by requiring every tweet to be at least 1,400 characters.

My new hobby: finding incomprehensible diagrams on office whiteboards and adding alarming conclusions to them