I am an undergrad at MIT studying Computer Science. I’m passionate about making AI and computing systems safe and efficient so they can benefit humanity at large.

Specifically, I wear two hats: the “science” hat where I explore model internals through mechanistic interpretability, and the “engineering” hat where I make GPUs go brrrrr.

Experiences

I’m currently a TA for 6.106 (6.172) Software Performance Engineering. I have worked on understanding neural network training dynamics in Max Tegmark’s group, compiler code generation using LLM with William Moses and neural scaling law at Epoch AI. See a list of my publications here.

News

• September 25, 2024: Paper “Algorithmic Progress in Language Models” is accepted as a poster at NeurIPS 2024!
• June 24, 2024: My paper “Survival of the Fittest Representation” is accepted to the “Mechanistic Interpretability Workshop at ICML”!

Contact

You can email me at carlguo[at]mit[dot]edu. You can also find me on GitHub, Google Scholar, LinkedIn, or Twitter. Submit anonynous feedback to me here.

My pronouns are he/him/his. I can speak English and Chinese Mandarin.

Misc

Before coming to the US for high school, I spent most of my life in Jinan, China, where I’m happy to call home.

In Winter 2023, I competed in MIT Battlecode and won 2nd place (read more here). I used to compete in college-level Mock Trial, for which I coded a website for automatic tournament tabulation.

In my free time, I like to make coffee, listen to music on records, and run.


Live the questions now. Perhaps you will then gradually, without noticing it, live along some distant day into the answer. —Rainer Maria Rilke