Hugo Vergnes

I train models from scratch and work out whether they actually hold up past benchmark numbers.

Right now that means video understanding at Apple. Before that I spent 4 years at Whiterabbit.ai on WRClear, a software-as-a-medical-device combining autonomous and assistive AI for screening mammography — developing and evaluating models across several clinical tasks, and leading statistical studies designed to demonstrate the device’s benefits to the FDA.

In the evenings I pretrain language models. The most recent one is little-lm: 3.8B parameters, trained from random weights to 0.384 CORE for $998 on rented GPUs. I wanted to watch language emerge from noise myself, and to learn the parts you only learn by starting from scratch.

I studied applied mathematics at École Polytechnique and statistics at Stanford. Outside work I train Judo (brown belt) and Brazilian Jiu-Jitsu (blue belt).