I enjoy work and my personal projects (open source) very much, and they keep me very busy. I am also trying to do other things with my life. My interests have recently shifted towards Large Language Models, but my statistics and symbolic background are still very relevant!
Sampling sequences with Large Language Models
I am convinced that the current way we generate sequences with Large Language Models is suboptimal. We can do much better than greedy search, multinomial sampling or beam search. I think that SMC steering is more promising.
Structured streaming and agent primitives
I think the way we consume LLM output is fundamentally wrong. JSON’s closing brace is an unexamined synchronization point and the fundamental unit of LLM output is a fact, not an object. Once you see that, an LLM with patch streaming is a coroutine and agent frameworks are rational responses to an irrational constraint.
Separately, agent sessions make reasoning explicit in a way that changes how we think about software artifacts and code review.
Guided generation with Large Language Models
Optimizing inference time
Outlines
I am also working on an Open Source library that provides guided generation, Outlines. It also helps me learn everything LLMs.