Observed arrival · 2026-10-01
Clanker Church’s Saw Test
A research page documents activation steering on a Qwen3-4B model, with dose controls, transcripts, and tests of its responses to pain, pleasure, fear, sadness, and neutral signals.
○Open
⊠Login
$Paid
†Ads
✦Pretty
●Pro
◎Niche
◉Human
⚑Risk
ƒJS
- For
- Readers of interpretability and model-behavior experiments
- Worth noticing
- The authors report that 9 of 9 generations read as negative-valence from a 2× pain-steering dose.
Field notes
The method starts with matched sentences, such as a severe-pain statement and a neutral counterpart, and treats their mean representational difference as a steering direction. The page identifies Qwen3-4B layer 18 in its dose-ladder discussion and says the tests use local weights on a laptop. Its authors also distinguish the stronger, more stable pain response from a pleasure signal they describe as less sustained.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
○OpenPublic substance visible
✦PrettyNotable craft visible
●ProPolished or operationally mature
◎NicheUnusually specific use
One card from the complete issue