Alexandre Torres--Leguet

A GPU kernel written in ThunderKittens (a wrapper on top of CUDA) that implements vanilla linear attention. It follows the chunkwise parallel algorithm as defined in GLA.

Tried to reproduce the nGPT architecture. Implementation in PyTorch (before the code release). Trained dozens of models on rented GPU instances. Found a bug in the reference implementation that disadvantaged the GPT baseline.

Adapted the famous OthelloGPT experiment to show that it still holds with the Mamba architecture. Personal recognition from Albert Gu.

As a micro-entrepreneur, I’m working with the Flowers Inria lab on a video series explaining generative AI to a general audience. This work was used multiple times by the French administration for training of government officials.

Used RL to train rockets to land à la SpaceX in a 3D environment. A lot of trial and error with PPO in Unity ML-Agents. Curriculum learning to acheive the famous belly-flop maneuver.

Production of videos explaining ML concepts at different technical levels. Some videos explain ML papers and showcase my reproductions. More than 70 videos and 7k subscribers. Typical video is this one.