end-of-text token between documents
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. 27 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 lights up on the hidden end-of-document marker that separates one web page from an unrelated one. Turning it up made the model's writing drift into odd, repetitive word loops, so the effect on text is hard to read.
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.
-
1, on the Mall in front of the National Air and Space Museum. The museum will also be hosting a scholarly conference on Civil War ballooning on June <|endoftext|>Most veteran peanut farmers in the Southeast will remember the devastating spread of tomato spotted wilt
-
be considered a planet, reigniting discussions.⏎In<|endoftext|>The Carnival season, or “Apokries” in Greek, is slowly coming
-
applications include improved communications technology, television displays and night-vision devices.⏎Tags: Lene Vestergaard Hau, physicist, spped of light.<|endoftext|>The Real
-
carried out for the samples with greenschist discontinuities, and found the characteristic of shear creep failure is closely related to the development of the structure surface.<|endoftext|>Japan’s Aerospace Exploration Agency (JAXA) is charting new territory in
-
on your feet. This can cause plantar fasciitis for the first time, or trigger a new bout once you have already healed.⏎If you know that<|endoftext|>CNN
-
, u, t)$.⏎<|endoftext|>While many students view summer as a time to forget about school, it remains an
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:iasmadowadowadowadowadowadowadowadowadowadowadowadowadowadowadowadowadowadowadowadowadowadowadowadowadowadowadowadowadowadowadowadowadowadowadowadowadowadowadowadowadowadowadowadowadowadowadow
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
’s article by Jeremy Rosen on Rav Lichtenstein is available from http://jeremyrosen.blogspot.co.uk/2015/08/rav-lichtenstein-zl.html⏎Your comment will be posted after it is approved.⏎
-
needed.⏎1. Children are starving as a result of famine. (is)⏎2. Where there is suffering that can be relieved, it should be relieved. (
-
against the Indian students. As a result, Subhash Ch<|endoftext|>BröstuligaOld Norse Dictionary - bröstuliga⏎Meaning of Old Norse word "bröstuliga" (or brǫstul
-
asti||ist, is (greek esti)|⏎|bandha (bind)||bind, bandage|⏎|iccha, isha (to wish)||wish|⏎|jan (produce offspring, born)||generate, genocide|⏎|k
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
- L25:239 strongly excites
- L20:563 excites
- L19:1174 excites
- L21:29 excites
- L19:509 excites
- L19:1207 suppresses
- L14:140 suppresses
- L22:65 suppresses
- L24:77 excites opening parenthesis after a name
- L18:76 suppresses blank spot before a number
- L19:124 suppresses
- L23:262 suppresses
- L25:222 suppresses
- L20:433 suppresses
- L23:80 excites blank slot before a year or number
- L19:1377 excites
- L20:1585 excites
- L21:321 excites
- L19:963 excites
- L22:2621 suppresses
- L22:416 excites commas in dates and number lists
- L23:10 excites
- L16:1421 suppresses
- L16:723 suppresses
Feeds later layers
None: this is the last layer.
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
- 三位一体
- 睁开
- 加盟
- 社团
- đạo
- decid
- tấn
- 木质
- 乙方
- 난
Suppresses
- ertain
- ascal
- amentals
- erval
- ousand
- irteen
- istency
- herited
- utions
- iversal