Scaling Compute on Context — Jack Morris, Engram

Very low training error on a document collection does not ensure useful answers about it. The talk description describes training a model on financial reports until its loss was extremely low, after which generation collapsed. Success on the training objective did not translate into usable behavior.

The proposed research direction is to make further computation produce progressively harder learning tasks from a fixed collection of material. A static set of generated exercises can eventually stop teaching the model anything new. Adapting those exercises as the model improves is presented as a possible route around that limit, with sustained improvement still the goal.

“Train a model directly on ten thousand financial reports and you can drive the loss to 0.00001. It knows the documents perfectly. Then you generate from it and it collapses.”

“Models are superb on Wikipedia, arXiv, and GitHub, and know nothing about your emails, your meetings, or your company.”

“Every approach shares a ceiling: you define a dataset, you train, and unless the model is underparameterized it eventually absorbs everything you made. A synthetic data wall, with none of pretraining's scaling behavior.”

“The property he wants is the one that made AlphaGo work, where getting better makes the training questions harder, so that adding compute keeps buying depth instead of flattening out.”