word-start 'm' pieces
A neuronOne of thousands of tiny dials inside each layer. It reads the text so far and produces a number, its activation, that nudges what the model writes next. in layerOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. 26 of Qwen3-0.6B. Its strongest snippets are very consistent: they mostly peak on the same token. It is featuredA neuron picked for the full treatment: steering outputs and a written explanation. About 3,000 of the 86,000 neurons are featured., so it has steering outputs.
What this neuron does
This neuron fires on a small 'm' that begins a word fragment, such as the 'm' in 'milder', 'mire' or 'mending'. Turning it up made the model repeat phrases or words like 'operating', which does not match the trigger.
Written by Claude from this neuron's activations and steering outputs. Confidence: medium. Think it is wrong? Flag it (opens a prefilled issue on GitHub).
Fires most on
The text windows where this neuron's activationThe number a neuron produces for a token. Large positive means it fired hard; negative means it pushed the other way. was highest. Amber marks positive activation; the outlined tokenA piece of text the model reads at once: often a word, sometimes part of a word or a punctuation mark. is the peak. Hover a token for its strength.
-
ándáil⏎Déantar do chuid sonraí íocaíochta a phróiseáil go slán. Ní dhéanann muid sonraí cártaí creidmheasa a stórá
-
packing them in oil or salt. Anchovy paste is also produced to use as an ingredient. Spanish boquerones are pickled in vinegar, which makes them milder.⏎Anchovies are notable in the culinary history of salt and<|endoftext|>M
-
1⏎Get out the microscope, because we’re going through this poem line-by-line.⏎No one has seen them made or heard them made,⏎But at spring mending-time we find them there.⏎- We get a bit distracted by all of
-
to return to normal⏎[weather, conditions] redevenir normal (e) → Even the weather has returned to normal.⏎(=regulation time) temps m réglementaire⏎at the end of normal time à l'issue du temps
-
vegetable stew, stuffed squash, green beans with rice, mouloukhiah, cauliflower stew.⏎4. Children meals preparation:⏎- Wash
-
will cluck like any hen,⏎And often to myself I think, why ye are just the sen!⏎– Singin’ blow, ye winds, in the mornin’,⏎– Blow, ye winds, high-ho!⏎– See
SteeringAdding a push along a neuron while the model writes, to see how its output changes. Turning it up or down.
Push this neuron up or down while the model writes, and watch what changes. The outputs were generated ahead of time, so nothing runs in your browser.
Here is something I have been thinking about lately: there is a t factor pean pean pean pean pean pean pean pean pean pean pean pean pean pean pean pean pean pean pean pean pean pe
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
even look like other software, making them virtually undetectable until it's too late.⏎The easiest way to fend off malware is to invest in antivirus and anti-malware software, where an on-access real-time scanner hooks deep into your operating
-
of the witnesses by which the facts are to be proved, or if there be written evidence, an abstract of such evidence; 5. The personal character of the witnesses should be mentioned; whether the moral character is good or bad
-
They suffered physical and emotional abuse, and may have been trapped in generations of servitude.⏎However, the Bible also includes some stories of how servants influenced their owners/masters towards good deeds. (As one example, see 2 Kings 5
-
his or her own initiative can make the appointment. The literal meaning of the phrase “Guardian ad Litem” is guardian for the lawsuit, or guardian for the purposes of
Connections
Show on mapWhich neuronsOne of thousands of tiny dials inside each layer. It reads the text so far and produces a number, its activation, that nudges what the model writes next. in earlier layersOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. feed this one, and which later ones it feeds. Amber links excite, blue links suppress; longer bars are stronger.
Fed by earlier layers
- L24:1190 excites
- L22:1138 suppresses
- L22:100 suppresses
- L17:230 excites
- L3:1101 suppresses
- L21:782 suppresses
- L18:553 excites
- L19:1531 excites
- L17:391 suppresses
- L17:241 suppresses
- L18:88 weakly excites
- L20:2288 weakly excites
- L19:381 weakly excites word-start 'acc' fragments
- L20:234 weakly suppresses
- L7:287 weakly suppresses
- L18:404 weakly excites 'get', 'go' and 'got' after a subject
- L20:212 weakly suppresses
- L7:342 weakly excites
- L19:334 weakly suppresses
- L8:2643 weakly excites
- L9:593 weakly suppresses
- L21:1679 weakly excites
- L1:1595 weakly suppresses
- L0:1010 weakly excites
Feeds later layers
- L27:1426 weakly excites
- L27:368 weakly excites
- L27:1319 weakly excites
- L27:1533 weakly suppresses
- L27:3070 weakly suppresses
- L27:2158 weakly suppresses
- L27:2241 weakly suppresses
- L27:2933 weakly suppresses the letter n opening a word
- L27:1097 weakly suppresses
- L27:2013 weakly excites
- L27:2282 weakly suppresses
- L27:1863 weakly suppresses
- L27:2810 weakly suppresses
- L27:1012 weakly excites
- L27:892 weakly excites
- L27:1976 weakly suppresses
- L27:939 weakly suppresses
- L27:480 weakly excites
- L27:1820 weakly suppresses
- L27:1826 weakly suppresses
- L27:2544 weakly excites
- L27:483 weakly excites
- L27:2142 weakly suppresses
- L27:2088 weakly excites
Activation histogramA chart of how often a neuron takes each activation value.
How often this neuron fires at each strength. Almost always it does close to nothing (the tall bars in the middle); the snippets above are the rare moments it fired hard, amber for positive and blue for negative.
Logit lensReading which output tokens a neuron pushes toward or away from, straight off its output weights, before the later layers have their say.
The words this neuron nudges the model toward (amber) or away from (blue) when it fires. A hint, not a label: many neurons work through later layersOne step in the stack the text passes through. Qwen3-0.6B has 28 layers, each with 3,072 neurons. instead.
Promotes
- 指
- RCS
- rake
- ima
- Organ
- räume
- ajas
- operating
- emaker
- recruits
Suppresses
- ưỡng
- angep
- PyErr
- .squareup
- 网络传播
- 监事会
- olas
- .Pixel
- Pew
- tang