Financial economics in the 1960s, and the music people

Trip 14 to 18 October 2024 · Written up 1 November 2024 · Last expanded 29 September 2026

October, and I was still crossing the Charles most mornings. The second half of the CMSA programme is where the argument I went there for finally turned up.

This looks a lot like financial economics in the 1960s

A day later I caught a parallel talk at Harvard by one of the workshop organizers, on how Turing’s work keeps shaping formal reasoning and mathematics, the second such conference for me. Researchers at Georgia Tech had landed on the same thought, as had the usual suspects out west. The sequel came after I had gone home: in December the CMSA invited a Georgia Tech project of mine to present, and Leyan Pan, my co-author on that first paper, delivered the overview. I watched it remotely, so it is not a pin on the map, but it was still my first academic collaboration since I left industry AI work.

The convergence itself is what interests me. Paul Samuelson said of Black-Scholes that after the rediscovery of Bachelier’s work it suddenly came to all the eager workers what the holy grail was, and the next step was the perfect formula to evaluate and price options. LLM formal reasoning feels like financial economics in the 60s, a flowering of approaches that are not yet really competing with each other, expanding every which way, and consolidation will come as some wither and others thrive. I would very much like to know which of the things on offer now is our Bachelier, already written down and waiting to be noticed.

At one of the Cambridge workshops the lecturer highlighted how hard it is to formalize mathematical material when the concepts are ill defined, and I think that is the real obstacle. I missed a CMSA session thanks to Memorial Drive traffic, François Charton and Geordie Williamson on how math and ML collaborations actually work, but got the long version from Charton at the hotel bar instead.

The music people were doing the more interesting version

Zhiyao Duan, who leads the University of Rochester’s Generative AI Music Lab, partitioned solutions by user type so cleanly that it shaped how I organized an MVP pitch for a humanities domain. In a second series of talks, Vivek H V of Naya Studio and Jeremy Joachim from Adobe’s Machine Intelligence group were the most thought provoking. They build workflow pipelines for creatives, where the craft is elevating someone’s own thoughts without overly directing them (generate a thousand artifacts with the primitive image tools and the regularity is hard to escape).

Once “best” stops being a single number, the end user cannot stay an afterthought, which is a lesson the music people seem to have absorbed earlier than most.

Who I met

• Zhiyao Duan, University of Rochester, Generative AI Music Lab. Partitioned solutions by user type so cleanly that it reshaped how an MVP pitch for a humanities domain was organized.
• Vivek H V, Naya Studio. Builds creative workflow pipelines that elevate a creative’s own thoughts without overly directing them.
• Jeremy Joachim, Adobe, Machine Intelligence group. The same theme from the other side, on why generative tools drift toward a regularity that is hard to escape.
• François Charton, affiliation not noted. Missed his CMSA talk with Geordie Williamson on how math and ML collaborations really work, then got the long version at the hotel bar.

Photographs

Links

Leyan Pan’s CMSA talk on transformers and SAT solving, and the Charton and Williamson CMSA talk on math and ML collaboration. To be added.

Open entry
This page is not finished, and is not meant to be. I add photographs, people and links as I dig them out, so it keeps growing after publication. Why the whole site works this way is on the colophon.