IndexMontréal / San Francisco
Building the computational layer for human dynamics.
Mathematics and philosophy at McGill, research at Mila. Working on AI in the space between people.
scroll00 — about
AI is not an intelligence problem, it is a relational problem.
I am in my fifth year of Honours Mathematics and Philosophy at McGill, and a researcher at Mila – Québec AI Institute. This autumn I am Researcher in Residence at The Residency in San Francisco.
My earlier research was about steering the dynamics of systems: interpretability for language models, weather models and agents. It has since converged on the relationship between people and AI — which environments, harnesses and protocols make the coupled system more intelligent, together.
Alongside the research I build places for people: dinners, salons, communities.
01 — inquiry
— noun · steering the dynamics of systems
Interpretability for language models, weather models, and agents.
Neural patterns, socially emergent collective intelligence, pro-sociality.
nowUnconventional forms of intelligence, substrate cohesion, AI in the spaces between.
nextCausal groundingMethodology firstHonest resultsOpen models
02 — work
GraphCast forecasts well, and nobody knows why. This audit edits the model’s internal features while it runs. A feature tied to the model’s own grid, not to any weather, turns out to drive forecast skill, and turning up convection and vorticity features pulls hurricane forecasts toward what happened. A plain statistical correction does about as well, and the paper says so.
Toward music chosen, and eventually generated, to drive a listener’s reward circuit as a way to reduce pain. So far a model can predict how a song looks to auditory cortex and which songs a person will like. The reward regions stay hard to predict, and the pain link is still a hypothesis.
Can a language model infer the mechanism behind cooperative behaviour, and does that mechanism have an identifiable direction in activation space? Sparse shift autoencoders and steering vectors in the iterated prisoner’s dilemma.
Two agents trained on a human trust game learn an internal map of the game that matches a social region of the human brain, and only after training. The brain-to-brain coupling I set out to compare against came out null, and is reported as null.
An instrument for what AI assistance does to your own reasoning. Paste a conversation with an assistant and it scores your half: Load, the share of the thinking you carried, and Rigor, the quality of it. No score without a receipt.
A night-time desert gathering. Twenty strangers dance out of time; you walk on and tap a rhythm; a circle ignites. A playable model of collective entrainment.
A cartography of the hidden metaphors, ontologies and reality gaps in multi-agent safety research — with a sixteen-type chart of assumption characters.
Carrying Goodfire’s Shape of Beliefs work over to a different open model, OLMo-2 in place of Llama, to test whether the same geometry of belief shows up there and whether steering along it still works.
Give it a place and your height and weight, and it ranks the nearby trees that are plausibly climbable, from city tree inventories, aerial imagery and street-level photos. It ranks; it never certifies a tree as safe.
03 — play
— for the fun of it
Two classes of groups defined by how well finite and matrix groups can approximate them, and why a 2020 result in quantum computing, MIP* = RE, which disproved Connes’ embedding conjecture, suggests that groups outside both classes should exist.
From Bolzano to Weierstrass: how a family of “pathological” functions forced analysis to become rigorous, and what counterexamples do for mathematics.
04 — words
essay · 19.07.2026
Building the computational layer for human dynamics.
“We become ourselves through each other.”Read on Substack
05 — pictures
“Attention is the rarest and purest form of generosity.”
— Simone Weil
06 — people
Coming soon.