Summary
Jack Morris, researcher and co-founder of Engram — a continual learning startup that launched the week of this talk — presents a framework for what he calls scaling compute on context. The core argument is that the standard three axes of AI scaling (data, parameters, compute) reduce to a single axis when you start from a pre-trained model and want to adapt it to private or specialized knowledge: compute. The talk positions Engram’s work as the pursuit of depth versus the breadth that pre-training delivers.
Morris uses Terence Tao as a framing device: current large language models can connect any public mathematical topic, but lack the subconscious intuition of a domain expert who has spent years inside a problem space. He argues that models have no principled mechanism to acquire knowledge after training — whether that is awareness of last night’s soccer scores or the idiosyncratic way your company’s partnership agreements are structured. He walks through several candidate approaches: naive next-token prediction fine-tuning on private corpora, synthetic data generation, and more compute-intensive study-time methods, evaluating each against the question of how to produce a model that genuinely knows a private dataset D.
The talk is deliberately high-level and conceptual rather than experimental, but it provides useful context for understanding the emerging continual learning startup landscape and how Engram differentiates from adjacent approaches like RAG, standard fine-tuning, and sleep-time compute proposals from other labs.
📺 Source: AI Engineer · Published August 12, 2026
🏷️ Format: Keynote Launch
