unclear: scattered mid-sentence tokens
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. 7 of Qwen3-0.6B. Its strongest snippets are mixed: they peak on different tokens. 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 grab bag of ordinary words in the middle of sentences, such as 'at', 'the' and 'is', with no clear shared meaning. Turning it up made the model get stuck repeating a single word like 'She' or 'The'.
Written by Claude from this neuron's activations and steering outputs. Confidence: low. 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.
-
station, municipal yard,⏎highway department, or school/library. There are some sites that⏎are operated by media such as newspapers, and a few that are at⏎ahead can help you locate just exactly where the station might⏎be
-
⏎Minerals can be identified by their hardness,⏎colour, the way they reflect light (lustre), the⏎way they break (cleavage and fracture), and⏎Mineral crystals fall into six crystal systems,⏎each of which is
-
that it⏎bridges them. Using a straw, blow air underneath⏎the paper. This causes a downward force on the paper, and it⏎moves down in the center.⏎By blowing the air between
-
Kansas, Texas, New Mexico, Montana, Alaska, and California; Canada=⏎Manitoba, Alberta, British Columbia; Mexico; Brazil; Poland; Denmark=Bornholm,⏎Greenland; Austria; Switzerland; Hungary;
-
shapes in this example, labelled 1, 2 and 3 as follows:⏎2222.2222 ..22.2... 11.222..3 ....2..33⏎
-
.⏎The surface is level or slightly undulating near the lake, but in the E.⏎part rough and mountainous, and more adapted to grazing. The soil is fertile,⏎particularly along the lake shore. Indian corn, wool, oats, and
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: the fact that, for example, the function f(x) = x^2 - 4x^2 + 6x^2 is the same as f(x) = 6x^2 - 4x^2.
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
historical State of the Union Address to more contemporary ones. In this Government instructional activity, students compare and contrast historical Presidential speeches to the ones they hear today. High...⏎9th - 12th Social Studies & History⏎Anticipating the
-
ons⏎together? How will you know you have found all the ways?⏎Swap the stars with the moons, using only knights' moves (as on a⏎chess board). What is the smallest number of moves possible?⏎Can you work
-
this⏎Dig into the entry’s quotation paragraphs, and you’ll see that the OED
-
02, 2012⏎Repairing bones with nanomaterials⏎(Nanowerk News) Around
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
- L4:156 strongly excites
- L4:224 excites
- L6:1374 excites
- L4:32 suppresses
- L5:1778 suppresses
- L1:2545 weakly excites
- L5:187 weakly suppresses
- L4:351 weakly suppresses
- L2:666 weakly suppresses
- L6:2586 weakly suppresses
- L4:138 weakly excites
- L6:191 weakly suppresses
- L5:238 weakly excites
- L1:1016 weakly suppresses
- L3:468 weakly suppresses
- L5:1240 weakly excites
- L2:673 weakly excites
- L6:203 weakly excites
- L6:2863 weakly suppresses
- L0:432 weakly suppresses
- L0:1403 weakly suppresses
- L1:1520 weakly excites
- L4:30 weakly excites
- L3:721 weakly excites
Feeds later layers
- L16:442 strongly suppresses
- L20:268 strongly excites
- L22:377 strongly excites
- L14:568 strongly excites
- L9:234 strongly suppresses line break inside hard-wrapped text
- L15:643 strongly excites
- L16:317 strongly excites
- L18:795 strongly excites
- L8:380 strongly suppresses
- L24:1619 strongly excites
- L26:1309 strongly excites
- L19:558 strongly excites
- L13:309 strongly excites
- L21:1164 strongly excites
- L25:2204 strongly excites
- L25:533 strongly excites first letter m or oct fragments
- L19:771 strongly excites
- L21:579 strongly excites
- L19:901 strongly excites
- L10:230 strongly excites
- L23:607 excites
- L16:795 excites
- L26:758 excites
- L27:947 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
- -⏎
- &⏎
- ,⏎
- ⏎
- ,⏎
- (⏎
- -⏎
- /⏎
- \⏎
- +⏎
Suppresses
- 县委书记
- ','',
- yat
- .'.$
- pewnością
- (*((
- ทำความสะอาด
- inely
- =").
- 联系我们