Simon Ouellette

Identifying the objects in an @arcprize grid is something I struggled to implement as a hand-crafted primitive, due to its "fuzzy" and contextual nature. I'm now doing instance segmentation using a transformer encoder to predict an embedding for each grid cell, trained on a…

In a way, what François Chollet did for the field of AI is to throw the @arcprize grenade into the "scaling is all you need"/"we already have AGI" crowd and watch as they all scrambled for the "oh no, I meant test-time compute scaling" and "reasoning LLM" exits. Then he saw that…

Free markets are a massively parallel, decentralized gradient‐descent procedure that discovers prices which clear supply and demand.

Intelligence is the ability that enables an agent to operate optimally in a non-ergodic environment. If the environment were ergodic, learning a compressed representation of past states would be enough.

Why did intelligence evolve? Because the universe is non-ergodic.

Upon waking up yesterday to the sudden news that o3 surpassed 85% on the public ARC-AGI benchmark, I had to update some fundamental things about my view of the current state of AI. Unlike all other previous approaches that claimed to reason but really weren't, this is a genuine…

I'm skeptical of Test-time training (TTT). I know it's behind the best scoring approaches on #arcprize at the moment. But I'm not convinced that it's because it allows the LLM to generalize structurally. My hypothesis is that TTT essentially does what I found my GridCoder 2024…

"Take, for example, the statistical independence between the identity and location of objects in the visual world – knowing what an object is does not tell us where it is. Strikingly, this statistical independence is reflected in the anatomical dissociation between the ventral…