François Grenier

From January to June 2026, at the CNRS (Lattice lab), I designed a new neural network architecture for deep language analysis. It is built on a new attention equation and a nested tokenization scheme for Transformers, trained and tested on the Jean Zay supercomputer.

The underlying problem is how to segment a continuous stream into meaningful units when the stream doesn’t come with its own words.

I study cognitive science, AI and mathematics at ENS Ulm (Cogmaster, the Master’s in Cognitive Science, M1 and M2), and I’m enrolled in the AI and Industrial Implementation track at HEC Paris. I entered both schools in 2023.

Computational and theoretical neuroscience

Synaptic plasticity and Hebbian learning rules, recurrent network dynamics, population coding and information theory, latent-variable and state-space models of neural activity, reinforcement learning and the role of dopamine.

Perception, attention and action

Psychophysics, efficient coding of natural stimuli, auditory perception, working memory, motor control and decision-making.

Machine learning for cognitive science

Bayesian inference and the predictive brain, neural time-series analysis, the interplay between deep learning and cognition, grounded in inferential and causal statistics.

Logic, linguistics and philosophy

Formal tools from logic and linguistics, and the philosophy of cognitive science.

I understand things by building them. I’ve rebuilt the modern ML stack from scratch (autograd, Transformers, diffusion, encoders, LoRA, inference and serving) and maintain a personal knowledge graph of more than 500 interconnected concepts across AI and mathematics. I’m also building a multi-GPU server dedicated to inference.

From March to June 2026, I was an AI Product Manager at one of France’s leading AI implementation firms. I led the implementation of complex AI solutions for public and private organizations, including the CEA (French Alternative Energies and Atomic Energy Commission), AS Saint-Étienne football club, and companies with revenues between €100 million and €500 million.

Before that, I advised Oxya Origin on AI and product strategy (2024–2025). I also co-founded and chaired HEC IA, an association of over 300 members with a 30-person board, for which I organized conference series and advanced technical training in generative AI.

More than 500 working notes on AI, connected in a graph: algorithmic concepts, neural architectures, mathematical foundations.

Open the graph in a new tab

Ancient Greek

A study tool for Ancient Greek texts, with the linguistic processing done at build time.

Russian

Tools I built to learn Russian.

I’m training as a pilot for humanitarian aviation, after working as a field assistant at a humanitarian relief center in Lviv in 2022.

Write to me at francoisemgrenier@gmail.com.