Matthew Berman discusses research testing whether a language model can infer properties of smaller neural networks from their weightsWeight space learning treats neural-network parameters as data from which a system can learn useful patterns. It can support analysis of model properties or methods that propose changes to another network. and write or edit weights for specific tasks. He walks through examples involving game-playing networks, simple image generators and language models.
Berman describes an edit that helped a small model avoid repetitive loops, while noting that fixing those loops did not necessarily improve its underlying mathematical abilityEvaluation is the systematic assessment of a system's behaviour or results against explicit criteria.. His speculation about self-improving AI goes beyond the specific experiments he presents.
Watch on YouTube
