Challenge: Hand coding weights for efficient sequence memorisation

Per the AI Alignment Forum, researchers have hand-coded the weights for neural networks that memorize token sequences, addressing a core question in AI interpretability. Their hand-coded models match the performance scaling of trained models, but still fall short by roughly a factor of ten at ninety percent accuracy. The team has issued a community challenge: find constructions that memorize more facts with the same number of parameters and close the gap to gradient descent solutions. This work contributes to a broader goal of understanding how large language models store factual information in their weights—a crucial step toward more interpretable and controllable AI systems.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton