Matt Langston, PhD

PhD physicist · Research Software Engineer
Applying experimental physics to understand how LLMs work
@Apple · formerly Stanford's @SLACLab

[Matt Langston, PhD @mattlangston]

Local Is All You Need.

Ran @karpathy's nanochat pipeline at home in 3 days on a DGX Spark Bundle ($8K once, $0 per run off solar). His speedrun does it in 3 hours on a rented 8xH100 for ~$48. Writeup: https://t.co/1JhRwHdpjH

[Matt Langston, PhD @mattlangston]

@AmandaAskell Get your pilots license. You’ll train in a Cessna. Stall training will make all flight simulators boring. Flying will put everything into perspective throughout your career no matter what you do.

[Matt Langston, PhD @mattlangston]

@ProfTomYeh @karpathy Pencil-and-paper backpropagation (derivatives + chain rule) for llm.c should maybe be step 0.

Matrix multiplication is important, but one won't understand the llm.c backprop matmul function calls without it.

[Matt Langston, PhD @mattlangston]

I wrote about my experience interviewing Claude Code for a software engineering position.
https://t.co/ob6whGutjg

[Matt Langston, PhD @mattlangston]

@pmddomingos Physicist here. I am no Einstein but loved every minute of grad school. I wouldn't be an RSE (best job in the world) without that knowledge, training and experience.

[Matt Langston, PhD @mattlangston]

@karpathy @ycombinator I've had so much fun and learned so much from your nanoGPT repo and videos - they're like @beneater's 6502 kit and videos for me.

I tinker with both all the time - what I was trying to convey.

[Matt Langston, PhD @mattlangston]

@karpathy 1980's PC era v2 for hobbyists. gpt2 is the new 6502.

[Matt Langston, PhD @mattlangston]

browsing macrumors