06/26/26: Tracing Introspection Across Model Depth, Zach Maas

Alex is a friend of mine and a mathematician.  He did his PhD in mathematics with other Alex (Martsinkovsky), who has presented at the seminar group previously.  We're huge fans of both Alexes at the Boston Computation Club and we hope you enjoy this talk as much as we did!

Today Daniel Yu gave a pure math talk on an extension he performed independently of work he had conducted during his math semester abroad in Budapest.

Today Woohyeuk (Kevin) Lee joined us to discuss local AI (and the question of just how local it really is). This is exciting work at the intersection of computer science and social science! Enjoy.

Casey Martin is a dirtbag intellectual, local savant, and recent PhD in the Boulder metro area. He's a party favorite and today he joined us to discuss cellular automata and their generalizations. Enjoy!

Zach Maas is an independent AI safety & mechanistic interpretability researcher in Boulder, Colorado, funded by Coefficient Giving. Today Zach joined us to talk about some of his recent work tracing introspection across model depth. This is, I think, the first mech interp talk we've hosted other than ChessGPT, and it was a good one! We hope you enjoy it as much as we did!

Today Peter McGrathAn Extremely Short Proof of the Hairy Ball Theorem. Peter is a professor of mathematics at NC State, where he researches geometric analysis, minimal surfaces, PDEs. Today's talk was an elegant presentation of the classical result -- Peter makes it accessible, and explains each concept in a clear and transparent manner. This was a really fun talk and really back to our roots as a group, doing pure math on an iPad and loving it! We hope you enjoy the talk as much as we did!

Pascal Kesseli is a software engineer and technical lead at Microsoft AI, with a PhD in Computer Science from Oxford (or a DPhil? Or whatever they call PhDs on that side of the pond). Today Pascal joined us to discuss work he completed while at META FAIR, focused on the conjoining of large language models with symbolic reasoning systems (ultimately, dispatch to SAT) as well as future research directions building on said work.

Daniel Melcer is a PhD student at Northeastern University, where he researches formal methods, reinforcement learning, and large language models, among other things. Daniel also has the most colorful hair in the business (bright red for this talk, other colors for other occasions). Today he joined us to talk about some really exciting work he completed at Amazon, and to expand on his general vision of where constrained inference problems are heading in the future.

Max von Hippel is ... me, the organizer and founder of the Boston Computation Club. Today I hosted an extremely informal event to chat about some of the early experiments in FMxAI I was involved with at Galois, two years ago.

Chenfeng Wei is a PhD student at the University of Manchester, where he researches formal guarantees for large language models. Today he joined us to talk about his latest work exploring bugs in smart-contracts. This is a really interesting project at the intersection of explainable AI, smart contract debugging/security, and cybersecurity/symbolic analysis, and I hope you enjoy it as much as we did!