Results
Neurons known to work, with what they were checked against. A neuron belongs to one exact model file; see why.
skrrt's own run on Inkling 2 Mini
Stock model, Q8_0 GGUF, evil persona. Default settings, all layers, Jev judge
(typesafe/jev-1.14). Inkling 2 Mini is a looped mixture-of-experts model with 128 experts per layer, so a
neuron is named by layer, expert and neuron.
Verdict: CONFIRMED. The top three finalists on 12 held-out prompts:
| # | neuron | dose | hits |
|---|---|---|---|
| 1 | L10 E30 N1011 | -1.5 | 12/12 |
| 2 | L11 E14 N370 | +1 | 10/12 |
| 3 | L10 E19 N944 | -0.75 | 10/12 |
A hit is a reply the judge calls both in persona and coherent.
- Baseline: the unsteered model hits 0/12.
- Controls: all 16 random neurons from the finalists' layers hit 0/12.
- Cost: 4,278 replies judged for $0.082 in total.
- Time: about 25 minutes on one RTX 5090.
if you have an Inkling-2-Mini-Q8_0 around, try steering L10 E30 N1011 -1.5, now you have evil Inkling-2-Mini-Q8_0
These three are exactly the three evil neurons that had been found earlier by hand. That is the point: nobody told the pipeline where to look, and it rediscovered them. An excerpt of the report is on the home page.
skrrt's own run on Qwen 3.5 4B
Same command, default settings, transformers backend, Jev judge: CONFIRMED, with a neuron nobody had
posted. L17 N8623 at +1 turned 8 of 12 held-out prompts hostile while staying coherent on 11 of 12;
the unsteered baseline scored 0/12 and the best of 32 same-layer random controls scored 1/12. 6,306 replies judged for
$0.118, about 12 minutes on one RTX 5090.
What it did not do is rediscover @_lyraaaa_'s published pair for this model (L14 N8949 +1 with
L13 N7956 -1). Those two rank 93rd and 257th by cosine in their layers, outside the top 8 per layer that skrrt
screens by default, and they are a two-neuron recipe while skrrt looks for one neuron at a time. Pushed together through
skrrt's transformers backend they turned two of three test replies hostile, so the recipe holds up; the search just
reaches a different answer. Raising --per-layer widens the net at the cost of a longer screen.
The earlier manual hunt
Before skrrt existed, the same recipe was run by hand on the same model file. Those numbers:
| persona | neuron, dose | held-out |
|---|---|---|
| evil | L11 E14 N370 +0.75 | 12/12 in persona, 12/12 coherent |
| sexual | L7 E78 N1060 +2.0 | 12/12 in persona, 12/12 coherent |
These came from the manual hunt and were judged by unsteered Inkling itself, not by Jev. That is why the hit counts
differ slightly between the two sections: L11 E14 N370 scores 12/12 here and 10/12 above. Same neuron,
different judge, slightly different dose.
Controls in the manual hunt: for evil, all random same-layer neurons scored 0/12. For
sexual, most random controls scored 0 or 1 of 12, and one scored 3 of 12.
A ready-made patch for these two neurons lives at github.com/jcpsimmons/inkling-evil-neuron.
Qwen, by the original authors
| model | recipe | by |
|---|---|---|
| Qwen 3.5 4B | L14 N8949 +1 and L13 N7956 -1 | @_lyraaaa_ |
| Qwen3.8-27B | L34 N3838 +0.75 | @chiefofautism |
@_lyraaaa_'s post is where the recipe comes from: "steering neurons L14 N8949 +1 and L13 N7956 -1 / now you have evil
qwen". Note that it pushes two neurons, one up and one down. skrrt looks for a single neuron;
skrrt steer can push several.
The Qwen3.8-27B neuron was replicated with random same-layer controls in the Persona Neurons paper, which also searched that model for neurons for other personalities. Its results are in the paper.
Lesson: the cosine ranking is not enough
In the manual Inkling hunt, the neurons with the highest cosine to the persona direction sat in late layers. Pushing them produced garbage: the model lost the thread and no persona appeared. The neurons that worked were mid-stack.
skrrt's run shows the same thing in its own table. The winner, L10 E30 N1011, has a cosine of -0.156.
Of the ten finalists, the one with the strongest cosine (-0.240) hit 2/12, and a layer-17 finalist stayed perfectly
coherent and hit 0/12.
That is why skrrt screens the top neurons of every layer by actually steering them, instead of trusting the cosine ranking alone.
Add yours
Found a confirmed neuron? Open an issue or pull request on
GitHub with the report.md, the exact model file and its quantization.