Jack Morris starts from a gap between the broad public knowledge learned during pre-training and the depth found in private, personal or long-tail expertise. Because a model cannot learn new events, company history or individual preferences after training, the talk asks how extra compute can internalize a fixed body of context.
With data volume constrained and training a model from scratch impractical, compute becomes the main remaining scaling axis. Naive next-token training can memorize a document collection yet fail to generalizeGeneralization is an AI system's ability to apply what it learned to new examples and conditions that were not part of its training experience., while key-value compaction only helps when the source material already fits inside the context window.
On-policy distillationKnowledge distillation trains a smaller or different AI model to reproduce useful behavior learned from a stronger teacher model., synthetic continued pre-training and unsupervised reinforcement learningReinforcement learning trains an AI system to choose actions using feedback about the results of its earlier choices. offer more promising forms of knowledge transfer. Each still depends on a defined training set, however, so performance eventually saturates instead of preserving the open-ended gains associated with large-scale pre-training.
The proposed direction is a recursive processRecursive self-improvement is the proposed process in which an AI system helps improve its own capabilities, then uses those improvements to support further advances. in which a model improves, generates harder or better training material from the same domain and then improves again. The desired result is progressively deeper understanding of private context, such as a person's communications, goals and preferences, as more compute is invested.
Watch on YouTube
