# @voooooogel — 2024-05-24

♥0 ↻0 · https://x.com/voooooogel/status/1794119869484679236

@karan4d not exactly, similar but different
- both are activation steering (inference time interventions)
- anthropic uses a pretrained SAE--they trained a big feature discovery model once
- control vectors are trained ad-hoc

tags: author:voooooogel, kind:tweet, thread-context, year:2024