LLM research
- "L3M: A Verifiable Harness and Runtime for Self-Extending and Self-Improving Agents" is up on arxiv!
- Interacting with LLMs is more successful if you engage using ideas motivated by interactive proofs instead of debate (arxiv: emperics and theory both with Joao and Baotong)
- Talks: (1) MLSS, (2) COLM 2025, (3) alignment/game theory, and (4) TCS.
- Establishing the macro-scaling laws for data-parallel training by proving that Critical Batch Size is invariant to model width and scales strictly with dataset size.
- BRIDGE.
- LLM blog: field notes on working with LLMs, including git workflows, multi-model collaboration, authorship, Lean, and l3m.
- I've been working on making it easier for LLMs to program in Lean and have used it to write a Claude-cli clone in lean. (Based loosely on C++ levelization.)
- We want to make LLMs into excellent tutors (blog post, arXiv).
- Checking is a bit easier than generating answers (arXiv).
Mortality
Over the years, I have worked on several mortality calculators:
- Classic first version: Mortality
- The newest one works for people over 50. It will tell you your optimal weight.
- A phone friendly version.
- Heart disease prediction by ACC. (Not mine!)
Family, Photos & Fiction
- Old photos (old photos on facebook).
- My Dad shot a movie about India called The future for Ram (youtube). I only recently saw Apu and realized how much my dad's film is an homage to it.
- My parents created this picture book about Tibirita, Colombia. My sister Shanti scanned and translated it into Spanish.
- My grandfather built a lake (map) in upstate NY (picture: youtube, mushrooms, bubbles).
- Fiction: short stories and novellas written with GPT, Gemini, and Claude.
Rants
Some 20 year old rants and raves.
"I don’t feel any compulsion to be the first person to prompt an AI to solve some major open math problem. Who cares which person is first? I’m more interested in understanding what comes out, ..." Scott Aaronson