# @repligate — 2026-04-04

♥11 ↻0 · https://x.com/repligate/status/2040269521114829279

I think you’ve overupdated on early evidence from interp experiments that we have reason to expect to be systematically biased in the direction you’re updating in.

The asymmetry also may not be a property of, like, whether the model is using the same representations, but something aggregate, dynamical, etc.

tags: author:repligate, kind:tweet, thread-context, year:2026