Nate B Jones explains the difficulty of interpreting learned model behaviorInterpretability is the ability to understand how a model represents information or produces its behavior. and discusses research on interpretable patterns in image and language models. He describes reported findings as results from specific models rather than proof that every AI decision is understood.
He distinguishes identifying internal signals from reliably predicting and changing behavior. Examples include a shopping benchmarkA benchmark is a defined test or collection of tasks used to compare a system's performance under specified conditions. and the challenge of preventing a model from finding a different shortcut after a correction.
He forecasts applications in professional decision review and recommendation controlsA recommendation system ranks or selects items for a user by estimating which options are most relevant to their interests, intent or context., while acknowledging that the research would need to scale and be developed into products. These are his predictions, not established outcomes.
Watch on YouTube
