linking words like 'further' and 'worse'
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. 25 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 responds to small connecting words that introduce a follow-up point, such as 'worse', 'better' and 'further'. Turning it up made the model keep producing the words 'still' and 'yet' over and over.
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.
-
literally hundreds of formats for graphical images. Although there are a few dozen formats publicly standardized, most are bitmapped formats -- especially hard to collaborate on -- and, worse, many incorporate data compression. Most of the rest require proprietary tools or operating systems
-
't complain (108).⏎7: Correct Usage 8 | An unlikely row has erupted in France over suggestions that the semicolon's days are numbered; worse, the growing influence of English is apparently to blame.⏎8: Correct Usage
-
Or better said: They should eat meat only.⏎Ferrets have a particular diet consisting
-
, it’s safe to say that there is no evidence that cell phones increase your chance of getting a brain tumor. Further studies will surely be conducted in the coming years but at this point, you’re
-
ould, who were collaborating with Frederick Law Olmsted in structures for Central Park, was soon eclipsed by the South range of the museum, by J. Cleaveland Cady, a robust exercise in rusticated brownstone neo-Roman
-
Cincinnati occurred in 1788 at the mouth of the Little Miami. Early entrepreneurs Israel Ludlow, Matthias Denman, and Robert Patterson purchased from John Cleves Symmes 800 acres of land and created the second settlement opposite
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: what is the more recent than the more recent? What is the more recent than the more recent than the more recent than... going on and on? What is the more recent than the more recent than the more recent than the more recent than
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
-edge techniques. Therefore, European industries involved in the development of sustainable seaweed aquaculture need to be supported.⏎Authors: Michèle Barbier, Bénédicte Charrier, Rita Araujo,
-
development of emotional intelligence in young children. As Robert Emmons, Ph.D. the world’s leading scientific expert on gratitude. a professor of psychology and founding editor-in-chief of The Journal of Positive Psychology, has discovered when people regularly engage in gratitude
-
thus ensuring tranquillity during much of his reign. In his final years, however, he faced internal disturbances, and he was ultimately killed in an attempted coup d’état.⏎Son of Setnakht (reigned 1190
-
He changed his approach immediately.⏎He was not a player in the October Revolution.⏎The Bolshevik revolution⏎Three interpretations of the October Revolution: popular uprising, coup d'etat
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
- L16:1733 weakly excites
- L18:650 weakly suppresses the word 'more' before a comparison
- L17:849 weakly excites
- L18:1454 weakly suppresses
- L8:14 weakly excites
- L13:2501 weakly suppresses
- L18:3065 weakly excites
- L17:2344 weakly suppresses
- L15:1259 weakly suppresses
- L13:1416 weakly suppresses
- L24:8 weakly suppresses
- L8:7 weakly excites
- L8:3 weakly excites
- L8:12 weakly excites
- L17:1900 weakly excites
- L4:92 weakly excites
- L16:248 weakly excites
- L22:535 weakly excites
- L1:1663 weakly excites
- L1:2133 weakly suppresses
- L4:1614 weakly suppresses
- L12:634 weakly excites
- L23:2405 weakly excites
- L8:44 weakly excites
Feeds later layers
- L27:1280 strongly excites
- L27:2009 weakly excites
- L26:2595 weakly excites
- L27:2489 weakly suppresses
- L27:644 weakly excites
- L27:1647 weakly suppresses
- L26:2690 weakly excites
- L26:814 weakly excites
- L27:1878 weakly suppresses
- L26:2996 weakly excites
- L27:572 weakly suppresses
- L27:2570 weakly suppresses 'con' starting a word piece
- L26:1262 weakly suppresses
- L26:1994 weakly suppresses
- L27:820 weakly suppresses
- L27:461 weakly suppresses
- L27:1254 weakly excites
- L27:1123 weakly excites
- L27:849 weakly suppresses
- L27:680 weakly suppresses
- L27:1132 weakly suppresses
- L27:3045 weakly suppresses capital W starting a word
- L26:1909 weakly suppresses
- L27:1341 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
- still
- still
- Still
- STILL
- Still
- 仍然
- yet
- 仍
- 依然
- 还是
Suppresses
- ave
- 第二节
- こと
- /uploads
- เฉพาะ
- inheritdoc
- enders
- andom
- (loop
- 楸