Howdy

I’m Max. I’m a Machine Learning Engineer and Researcher, first working at KUNGFU.AI applying ML as an engineer, then cohere.for.ai doing research, and now at MosaicML. I published a paper on data pruning (the longer pre-print here) for LLM pretraining, and my other work has covered LLM evaluations, computer vision, audio, and more.

I enjoy backpacking where I can forget about the anthropocene, have spent a whopping <REDACTED> hours playing Dota 2 (go zai, go rtz, go cr1t), lift, surf, and spend as much time outside as I can. In Austin I played ultimate frisbee on Riverside and loved hanging out in my backyard with my cat, Callie. I grew up in Austin, went to school in Los Angeles, moved back home, and now live in San Francisco.

poor attempts at professionalism here

worst takes here

research only here

code here

my str8 up email (remove the hightech captcha): m m a r i o n 5 3 8 @ < r o b o t s g o a w a y > @ g m a i l

I was recently briefly featured in this Scientific American article on LLMs and their possibilities for religious applications and this Wall Street Journal article on Noah Lyles and the rise nerds.

Here’s a list of the cooler projects I’ve worked on, during which my role always fell under “make the neural network work”:

a preprint exploring various methods of judging language data as low or high quality (paper link)
LLM eval for SOTA LLMs currently in production
an algorithm to judge the price of collectibles like Pokemon cards, baseball cards, or rare coins
a computer vision project identifying and pricing exterior damage on cars from cell phone photos
a model for processing ambient rainforest noise to find gunshots, chainsaws, and vehicles indicating poaching and logging
a system to scan insurance corpora to extract helpful sentences like “drug X is covered by plan Y under situation Z, but not situation A”
a CTR (Click Through Rate) predictor for crowd groaning noises ads
a FOD detector for runway radar scans (detects debris on your runway that’d destroy your fancy plane engines)
A number of “information extractors” for large datastores of pdfs
the algorithm that judges SXSW speaker applications alongside human judges
and more! that are less interesting and/or slightly secret