'to' after come or bring
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. 7 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 'came to be', 'come to see', and 'brought it to my attention', where a verb of arriving or bringing is followed by 'to'. Its negative side is 'to' in 'change to' or 'by' in 'replaced by', which are different uses. Turning it up gave text with no clear link to these phrases.
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.
-
book, The Price of Privilege, Marilyn Levine tries to make sense of this trend: “When parents place an excessively high value on outstanding performance, children come to see anything less than perfection as failure. While most kids hang in there and try
-
that their soldiers were well-supplied with ample shields.⏎Shields were used by both the cavalrymen and infantrymen in a Carolingian army. This came to be due to the fact that Carolingian shields were very inexpensive and even common
-
complexity of individuals relationships to the world they live in and the effects on individual health of pollution, poverty, and unhealthy social norms and values are ignored. People come to construe healthy behavior in terms of dependency on a medical industry that constantly invents
-
periodically boosted the space station's orbit, and in fact helped the station dodge a passing piece of Russian space junk last month.⏎But all good things must come to an end: Unlike the Italian-built space cargo modules that are carried back and forth inside
-
Quality Control (QC) and Testing. Though they’re closely related, they are, ultimately, different.⏎If you work in the IT industry, you’ve probably come across them. You’ve also probably noticed that many executives – and customers – don
-
Believe it or not, the sauce was brought to France by an Italian chef called Catherina de Medici in the 1
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: a more detailed and comprehensive plan of a 10-day plan for a 10-week plan. So I will just focus on the 10-day plan. So I will just focus on the 10-day plan for a
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
: ‘do not’ would change to ‘don’t’⏎That’s it, you’ve got it…. Now write
-
follow this rule (as in soliloquies).⏎Add -s to most nouns ending in -f.⏎Exceptions: In some cases, change the -f to -v and add -es⏎In compound words, make the main word plural
-
.⏎- If stuck for choice, pick the best of a bad bunch;⏎Long commutes to, from or during work can mean that sit-down meals are replaced by unhealthy foods on the go. It also means that you are constantly exposed to food
-
remove the carriage return and replace it with a space, and then break your line of text at a point that is convenient
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
- L6:984 suppresses
- L2:750 excites
- L6:1702 suppresses "not" or "instead" contrasting a correction
- L3:1152 suppresses
- L5:1671 weakly excites
- L4:1531 weakly suppresses
- L4:2177 weakly excites
- L4:328 weakly suppresses 'to' after 'so as' and infinitive cues
- L1:2630 weakly suppresses
- L5:548 weakly excites
- L1:2525 weakly excites
- L5:3009 weakly excites 'to' after a verb or noun
- L1:906 weakly suppresses
- L4:2961 weakly excites
- L5:2867 weakly excites
- L4:2111 weakly excites
- L6:1974 weakly excites
- L4:123 weakly suppresses
- L4:1847 weakly suppresses
- L5:770 weakly suppresses
- L6:2451 weakly suppresses
- L5:1802 weakly suppresses
- L4:1236 weakly excites
- L5:2057 weakly suppresses
Feeds later layers
- L9:723 suppresses
- L16:937 suppresses
- L11:1895 suppresses
- L17:429 suppresses
- L17:2007 suppresses
- L13:383 excites
- L14:785 suppresses
- L23:477 weakly suppresses
- L20:1465 weakly suppresses
- L9:2301 weakly excites
- L16:2763 weakly excites
- L15:2958 weakly suppresses
- L15:1238 weakly excites 'less' before a comparative or negative effect
- L16:804 weakly excites
- L8:909 weakly excites
- L15:428 weakly suppresses
- L15:2981 weakly suppresses
- L13:2082 weakly suppresses
- L18:2919 weakly suppresses
- L9:3000 weakly excites
- L8:2298 weakly suppresses
- L25:2288 weakly excites
- L17:2598 weakly suppresses
- L12:2168 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
- !.⏎⏎
- 头发
- evaluation
- getData
- ucket
- 大发
- braco
- 的灵魂
- 似的
- kitchens
Suppresses
- 另一
- 新一代
- =current
- دد
- 明年
- 二线
- ymi
- conserve
- ampa
- yms