Observed arrival · 2026-10-29
Poor Things’ neuron laboratory
A public lab record of experiments steering individual Qwen3.5-4B neurons and scoring the model’s responses.
○Open
⊠Login
$Paid
†Ads
✦Pretty
●Pro
◎Niche
◉Human
⚑Risk
ƒJS
- For
- Readers studying activation steering and model behavior
- Worth noticing
- Each specimen uses a signed dose on one MLP neuron or a pair, compared with an unsteered control.
Field notes
The procedure specifies how intervention size is normalized: dose is scaled by the layer’s median residual norm and the neuron’s down-projection weight norm. The accession ledger reports 2,772 settings and 92,960 stored replies, while the page says the rankings change as more specimens are run; it explicitly treats specimen names as working labels rather than evidence of a model’s beliefs or consciousness.
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