Author here.<p>This post shows “concept algebra” on language model: inject, suppress, and compose human-understandable concepts at inference time (no retraining, no prompt engineering).<p>There’s an interactive demo on the post.<p>Would love feedback on:
(1) what steering tasks you’d benchmark,
(2) failure cases you’d want to see,
(3) whether this kind of compositional control is useful in real products.<p>Related: <a href="https://news.ycombinator.com/item?id=47131225">https://news.ycombinator.com/item?id=47131225</a>