'to' after 'so as' and infinitive cues
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. 4 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' in phrases like 'so as to' or 'a party to', where a purpose or infinitive is about to follow. Turning it up gave no clear theme in the output.
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.
-
(cause to go off) "fire a gun"; "fire a bullet"⏎- S: (v) fire (bake in a kiln so as to harden) "fire pottery
-
educational computers are so busy trying to maintain their market share that they failed to see the logic of working to the same basic voltage in selecting their architecture, so as to be more compatible with one another. If they were compatible
-
get to normal subduing the demand. The government starts to deepen the pocket by increasing taxes, so as to benefit from the economic boom.⏎In other words, the economy is too complex to
-
the need for any investigation into the likelihood or suspicion of bias. The most common form of imputed bias is where the judge or the decision-maker is a party to a suit – in other words they are actually involved in the dispute that appears before
-
for show-up. He or she can be required to speak, stand in a particular pose, and wear appropriate clothing.⏎Suspects can invoke the Sixth Amendment right to counsel during lineups and showups held after they have been arraigned or indicted
-
capital Kt: with BK, BKK >0.⏎Consumers save in the form of capital; the representative consumer chooses an optimal consumption plan, so as to maximize⏎where c is consumption per capita and r is the rate of time preference
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 to get some sleep. I have to get to the point of sleep. It is not a thing to think that sleep, and to get to the point of sleep. I have to get to the point of sleep. It is
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
there's great hope⏎If we our eyes do only ope.⏎With joy they will often glisten,⏎If to truth one doth but listen ;⏎But people often turn deaf ear⏎And what
-
that compensates for the impairment of functional networks caused by brain damage. A reserve that is constituted from childhood and that we can continue to feed throughout life.⏎And to stimulate the brain there is no mystery: You have to use it. Those who
-
was disclosed.⏎She let the cat out of her bag of verse she almost proposed to her hero in rhyme.George Meredith: The Egotist, iii.⏎To bell the cat. (See page 119, B
-
set forth in the “preamble,” or opening clause, of the Constitution of the United States;⏎1. “To form a more perfect union;”⏎2. “To establish justice;”⏎3
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
- L3:1152 suppresses
- L2:750 excites
- L3:1859 suppresses
- L1:906 suppresses
- L3:2705 suppresses
- L3:2947 suppresses
- L3:59 excites
- L3:1573 suppresses
- L1:2296 excites
- L3:1016 weakly excites
- L0:1551 weakly suppresses
- L2:1036 weakly excites
- L3:691 weakly excites
- L3:2818 weakly excites
- L3:456 weakly suppresses
- L2:953 weakly suppresses
- L3:50 weakly excites
- L0:1301 weakly suppresses
- L0:178 weakly suppresses
- L3:115 weakly suppresses
- L2:1899 weakly excites
- L2:126 weakly excites
- L3:311 weakly suppresses
- L2:2866 weakly suppresses
Feeds later layers
- L19:858 strongly suppresses
- L5:2251 suppresses
- L8:1627 suppresses
- L18:1001 suppresses
- L16:1168 suppresses 'to' after verbs of change
- L10:132 suppresses
- L9:723 suppresses
- L5:2538 suppresses
- L9:1890 suppresses
- L17:1755 suppresses
- L17:1297 suppresses
- L9:2267 suppresses 'with' after exchange nouns
- L12:2084 suppresses
- L11:437 suppresses
- L5:2026 suppresses "to" after made or need-type verbs
- L27:2960 suppresses
- L19:2395 suppresses
- L6:2686 suppresses
- L6:2306 weakly suppresses
- L6:2899 weakly suppresses
- L5:363 weakly excites
- L7:1014 weakly suppresses
- L17:1843 weakly suppresses
- L14:2972 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
- bí
- ifter
- ou
- istinguished
- pip
- [('
- \$
- (\$
- ited
- Petr
Suppresses
- 阗
- hn
- Neighbors
- gether
- (rawValue
- ffee
- MMI
- )object
- 澍
- ollo