'to' after verbs like continue or help
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. 8 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 the word 'to' when it follows verbs such as 'continue', 'help', or 'set', as in 'continue to reverberate'. It goes the other way on 'been' after 'has' or 'had'. Turning it up gave unrelated text, so the steering results do not show the pattern.
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.
-
⏎(Nanowerk News) Around half a million people in the EU need treatment for serious skeletal defects every year. New research on bone tissue and repair promises to overcome inherent problems when a large bone graft is required.⏎Therapy of large
-
ricated. Regular maintenance helps detect wear and tear early, preventing small issues from escalating into major repairs. This proactive approach to maintenance extends the life of the HVAC
-
United States. The meeting took place on July 19-20, 1848, at Seneca Falls, New York. It helped to launch the U.S. campaign for women’s suffrage, or the right to
-
<|endoftext|>These days computer-generated fractal patterns are everywhere, from squiggly designs on computer art posters to illustrations in the most serious of physics journals. Interest continues to
-
natural resources consciously and sustainably.⏎The shocking images and news of the disastrous flooding in the Eifel, North Rhine-Westphalia and Bavaria continue to reverberate. They have once again raised our awareness of the fact that climate
-
: We have already seen a global reduction in air pollution due to lockdown and less travel. If we still use our cars less after the pandemic, this could continue to make a huge difference.⏎But, how much of a difference are we talking?⏎
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: I have been thinking about lately about the value of my life, and I am going to make a decision. So far, this is the first time I have been thinking about it. I have been thinking about it. So I am going
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
January or February with one-third of the tree's total annual nitrogen requirements. The amount depends on the tree's size and how many years it has been in the ground. A mature tree that has been planted 6 or more years
-
decide what other people can do with it. Copyright laws make it easier for authors to make money by selling their works. Because of copyright, a work can only be copied if the owner of the copyright gives permission.⏎When someone copies or edits a
-
happy or sad in the garden? Why?⏎- Who had the giant been visiting?⏎- How long had the giant been away?⏎- In your own words describe what the giant did when he arrived at
-
sensitive,’ ‘too emotional,’ and that their behaviors are either ‘too dramatic’ or ‘too timid.’⏎People who feel more deeply and intensely than others are more aware of subtleties; their brain processes information and reflects on it more deeply
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
- L7:681 excites 'from' after prevent or stop
- L6:914 excites
- L6:439 suppresses
- L6:2747 suppresses
- L6:2359 excites
- L6:1133 excites "to" before an infinitive verb
- L4:518 suppresses
- L6:1621 suppresses
- L4:113 suppresses
- L6:2252 excites
- L6:1269 excites
- L7:6 weakly suppresses
- L4:521 weakly suppresses
- L6:1880 weakly excites "by" introducing a deadline or means
- L6:1103 weakly suppresses
- L6:2297 weakly excites
- L4:1394 weakly suppresses
- L2:1979 weakly excites
- L6:95 weakly suppresses
- L6:2091 weakly suppresses
- L6:1751 weakly excites
- L4:2216 weakly excites
- L3:755 weakly excites
- L5:1505 weakly suppresses
Feeds later layers
- L9:459 suppresses
- L12:2176 suppresses 'of' after a place or role noun
- L20:113 excites
- L9:222 suppresses
- L9:188 suppresses 'of' in 'National Institute of' names
- L19:1582 suppresses "of" after an abstract noun
- L13:1100 suppresses
- L18:1645 weakly excites
- L10:459 weakly suppresses
- L21:215 weakly excites
- L17:1225 weakly suppresses
- L9:52 weakly excites
- L24:85 weakly suppresses
- L17:2057 weakly suppresses
- L16:987 weakly suppresses
- L9:335 weakly suppresses 'of' after nouns like anniversary or onset
- L15:317 weakly excites
- L9:841 weakly suppresses
- L15:696 weakly suppresses
- L14:672 weakly suppresses
- L14:2275 weakly suppresses
- L18:1253 weakly suppresses
- L20:480 weakly suppresses
- L13:700 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
- �能
- ensor
- 不容
- abric
- emain
- 不留
- ource
- roller
- ьер
- Uri
Suppresses
- 打扫
- gang
- ],'
- movement
- /documents
- (TAG
- /fwlink
- ],"
- (!(
- %; ⏎