Jack Morris: Stuffing Context is not Memory, Updating Weights is

Jack Morris, a researcher presenting at the AI Engineer conference, challenges a common assumption in AI system design: that putting data into context is the same as a model ‘knowing’ something. His talk, titled ‘Stuffing Context is not Memory, Updating Weights is,’ argues for a third approach to knowledge injection — training information directly into model weights — and explains why this matters for tasks that RAG and long-context windows struggle with.

Morris systematically walks through the tradeoffs of each approach. Context stuffing is simple but slow and expensive at scale — pasting 80 pages into Claude, he notes, slows inference by roughly 10x. RAG is efficient but breaks down for complex, interrelated documents where retrieval precision is insufficient. His proposed alternative: fine-tuning models on domain-specific data so that knowledge becomes part of the model’s parameters, making it fast, always accessible, and free from retrieval errors.

The talk covers practical implementation questions — SFT vs. RL training, data augmentation strategies, and why architectural choices may become relevant again as personal or organizational models get continuously updated. Morris illustrates with a concrete experiment fine-tuning on 3M financial reports, honestly discussing both what next-token-prediction training achieves and where it falls short. This is an accessible but technically substantive talk for engineers building systems that need models to reliably know domain-specific, proprietary, or longitudinal information.

📺 Source: AI Engineer · Published December 29, 2025
🏷️ Format: Deep Dive