Aiman Shabib

I started in the engine. Unity, C#, shipped VR titles, mixed-reality prototypes built in two-day hackathons. That world taught me something I keep coming back to: every frame is a deadline, every abstraction has a cost. Real-time systems are honest in a way ML benchmarks usually aren't.

Most of my time now goes into research. The main line is Parallax, where I am poking at what happens if you stop treating cognition as a token-clocked autoregressive function and instead let it settle - more like an attractor than a sampler. Around that I keep a notebook of smaller side experiments. A dendritic block without an activation function. Evolutionary agents in Unity. An internal Pi project. A lot of them will probably go nowhere. Some have already taught me things I would not have figured out by reading papers.

What I actually care about is the through-line. Why does some compute give rise to behavior that looks like understanding, and other compute does not. I read this less as engineering and more as physics - the interesting problems live in the geometry of representations, in the shape of the bottleneck, in what gets baked into the unit before any weights are learned.