News for hackers. Stuff that matters.

A comment on Show HN: Steerling-8B, a language model that can explain any token it generates.

Re:Show HN: Steerling-8B, a language model that can explain ... by andy12_ on Tuesday February 24, 2026 @04:53AM (#47135060)
This seems really interesting. While Anthropic tried to use dictionary learning over an existing model to try to extract concepts, this almost feels like training the model alongside the dictionary itself (or rather, the model and the dictionary are intertwined).
[ Parent | on HN ]
Re:Show HN: Steerling-8B, a language model that can explain ... by adebayoj on Tuesday February 24, 2026 @05:00AM (#47135101)
You are exactly right, it is guiding the model, during training, with concepts and the dictionary. This is important because dictionary learning for interpretability (post hoc) is not currently reliable: https://www.arxiv.org/abs/2602.14111
[ Parent | on HN ]