1 pointby luulinh90s9 hours ago1 comment
  • giang_at_glai3 hours ago
    Author here.

    This post shows “concept algebra” on language model: inject, suppress, and compose human-understandable concepts at inference time (no retraining, no prompt engineering).

    There’s an interactive demo on the post.

    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.

    Related: https://news.ycombinator.com/item?id=47131225