end of document after a full paragraph
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. 12 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 responds strongly to the hidden end-of-text marker when it comes after a complete paragraph, so it seems to sense that one piece of writing has finished. Turning it up did not give a clear theme.
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.
-
their . place. So, the students, teachers and other professionals finally had to resume their studies and jobs in government institutions like schools, colleges, courts etc.<|endoftext|>I hope you are all well and that you enjoyed some of your weekend in the
-
own how to get there in the same way they landed the jet five years before.⏎So, there are lessons here:⏎- Without leadership from the top, bottom<|endoftext|>Updated: Jun 1, 2020⏎Jaundice is
-
–6. Warsaw, 1877–82.⏎Pamiqtki Soplicy. Warsaw, 1961.<|endoftext|>Translate this page into:⏎Approach to family assessment and intervention⏎All India Institute
-
i>⏎;; Now insert one character at the beginning of the buffer: (goto-char (point-min)) ⇒ 1 (insert "Q") ⇒ nil⏎<|endoftext|>A bead is a small bauble or jewel with a hole through its middle for
-
to break down carbohydrates, proteins and fats.⏎You may have an idea, but there are many kinds of digestive enzymes that break down different kinds of food items.<|endoftext|>Chinese architecture started from ancient cave-dwellings, that are still inhabited on the
-
editing nightmare and when the sorting is finished, it will put all of the blank spaces at the top of the list, leaving the data single-spaced anyway.<|endoftext|>A hydrogen atom falling on a neutron star surface from infinity releases 200
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: X. A. So I Have. A. So A. Therefore. So A. Therefore. A. So A. Therefore. So A. Therefore. So A. Therefore. So A. Therefore. So A. Therefore. So
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Countess of Rosebery, died, at Dalmeny House, in Scotland, at the
-
systems.⏎What is Lyra's demon?⏎lyra's demon, pantalaimon /ˌpæntəˈlaɪmən/, is his most beloved companion, whom he calls “
-
10101100.0001000x.xxxxxxxx.xxxxxxxx SUBNET IP: 10101100.0001000x.xyyyyyyy.y
-
Holiday Crowd At The Zoo⏎- The Child Is Father Of The Man (Es⏎Past Perfect Tense Interrogative⏎Had + subject + verb 3rd from + object.⏎Past Perfect Interrogative
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
- L10:174 weakly excites
- L9:1423 weakly excites
- L4:1697 weakly excites
- L0:1414 weakly excites
- L8:1567 weakly suppresses
- L7:2237 weakly excites
- L8:1290 weakly suppresses
- L4:542 weakly suppresses
- L9:428 weakly excites split word pieces in names and years
- L10:229 weakly excites
- L1:2833 weakly excites
- L1:2458 weakly excites
- L6:2296 weakly excites
- L10:348 weakly suppresses
- L5:2422 weakly suppresses
- L4:2470 weakly suppresses
- L9:1531 weakly excites
- L5:1693 weakly suppresses
- L10:2654 weakly suppresses
- L5:1062 weakly suppresses
- L3:2490 weakly excites
- L0:555 weakly suppresses
- L9:1589 weakly suppresses
- L7:2364 weakly suppresses
Feeds later layers
- L17:712 weakly excites
- L19:809 weakly suppresses
- L14:2525 weakly excites
- L27:1696 weakly suppresses
- L14:177 weakly excites
- L22:2753 weakly suppresses
- L17:32 weakly excites end of document, boilerplate footers
- L18:783 weakly suppresses
- L20:750 weakly suppresses unclear: chinese academic text, 'when' loops
- L21:2151 weakly suppresses
- L26:2780 weakly suppresses
- L18:1679 weakly suppresses
- L15:1601 weakly suppresses
- L21:250 weakly excites
- L15:2463 weakly suppresses
- L23:2922 weakly suppresses
- L24:174 weakly excites
- L15:448 weakly suppresses
- L18:119 weakly excites
- L14:469 weakly excites
- L13:1001 weakly suppresses
- L23:734 weakly excites
- L17:57 weakly excites
- L14:2152 weakly suppresses
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
- !
- ax
- ady
- ax
- (ax
- asjon
- :
- ę
- ?
- [
Suppresses
- ��
- '\\'
- Serviço
- "\\"
- 经营模式
- TURE
- (','
- 嘉年
- .Resume
- hire