Engram Is Now Public: A Framework for the Execution Substrate of Modern AI Models

Today I’m releasing Engram, a framework that names and formalizes the layer where modern AI models are actually realized during inference — the execution substrate.

For years, AI research has treated inference as a neutral evaluation of a mathematical function. But anyone who has watched a runtime update change a model’s reasoning without touching the weights knows this assumption no longer holds. The trained model is not the model we interact with. What we observe is the realized model, shaped by precision formats, memory hierarchy, scheduling, topology, and numerical stability.

Engram gives this layer a name, a vocabulary, and a research program.

The paper is now publicly available:


This release is part of a broader effort to map the climate of the Second Silicon Winter — where memory, execution, and resource constraints shape cognition as much as weights and architectures do. Engram focuses on the execution‑level component of that climate: the substrate through which reasoning must physically pass.

If you’re interested in inference drift, reproducibility gaps, execution‑aware cognition, or the future of ML systems, I think you’ll find Engram a useful lens.

More to come soon.