'they' or 'we' after where or who
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. 26 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 a pronoun like "they" or "we" that comes right after words like "where", "who" or "if", as in "who we are" or "where they started". Turned up, the output drifted into unrelated text about news and reflections.
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.
-
positive and negative, which sum to zero.⏎The positive cycle may move charge carriers, such as electrons. But then the negative cycle pulls the charges back to where they started. To reliably control the quantum information, an asymmetric light wave is needed.⏎
-
the starting point illustrated in their body rotation. Most of the participants could use information about where they were in the room just a few moments before to tell them where they were in relation to the starting point.⏎Further research into where the “noise”
-
included this in Genesis 3<|endoftext|>Sujata Gupta for Nature – Brain food: Clever eating:⏎Our ancestors’ consumption of meat is what made us who we are.⏎
-
continue building out ground stations as it gets more approvals from the regulatory bodies to do so in the United States and Canada.<|endoftext|>Our brain is what makes us who we are, is it not? All aspects of our life are stored into this one
-
pointier and whiter. So the size and other teeth features should help you determine its age.⏎Also, check closely if the permanent teeth are white. If they still are, then your cat is probably around 16
-
chalices, and other spoils of the Church might be as much more. These movables were before night taken to the King’s treasury—a sign that they will not be restored.”1⏎The Spanish Chronicle also reported Thomas Cromwell’s
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: howbeitbeitartreambeitreambeitreambeitreambeitreambeitreambeitreambeitreambeitreambeitreambeitreambeitreambeitreambeitreambeitreambeitreambeitreambeitreambeitreambeitreambeitreambeitreambeit
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
evaluations about jesuitic missions among the Guaranis - Relação abbreviada da republica, que os religiosos jesuitas estabelecerão no Paraguai, by Marquis of Pombal, Candide
-
a masculine, plural noun so the masculine plural direct object pronoun los is used to replace it in the second sentence.Son muy buenos hijos. No los cambiaría.⏎They are good
-
(Bot.) A smooth-skinned variety of peach. [1913 Webster] Spanish nectarine, the plumlike fruit of the West Indian tree Chrysobalanus Icaco; -- also called cocoa plum. It is
-
2018, Schuldt et al. 2020). Under these circumstances, tree´s survival
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
- L23:625 strongly suppresses
- L22:309 strongly excites
- L24:120 strongly suppresses
- L24:241 strongly suppresses
- L23:176 strongly suppresses
- L24:600 strongly excites
- L20:1423 strongly suppresses
- L23:287 strongly suppresses 'are' after a plural subject
- L21:468 strongly excites
- L24:2880 strongly excites
- L19:1317 strongly excites
- L25:2055 strongly suppresses
- L25:2082 strongly suppresses first-person 'I' starting a clause
- L23:1210 strongly suppresses
- L19:1422 strongly excites
- L24:401 strongly suppresses 'it' referring back to a noun
- L21:2936 strongly suppresses
- L18:444 suppresses
- L24:2161 excites
- L20:2453 excites blank slot before a date
- L12:642 excites
- L25:586 suppresses
- L16:1317 excites
- L2:149 suppresses
Feeds later layers
- L27:500 strongly suppresses
- L27:1256 strongly suppresses
- L27:321 suppresses
- L27:650 excites
- L27:461 suppresses
- L27:1040 suppresses
- L27:804 weakly suppresses
- L27:2212 weakly suppresses
- L27:2329 weakly suppresses
- L27:1271 weakly suppresses
- L27:1799 weakly suppresses
- L27:3047 weakly suppresses
- L27:1911 weakly suppresses
- L27:168 weakly suppresses
- L27:995 weakly suppresses
- L27:2196 weakly suppresses
- L27:1740 weakly suppresses
- L27:3032 weakly suppresses 're' prefix meaning again
- L27:485 weakly suppresses
- L27:1357 weakly excites
- L27:1937 weakly suppresses
- L27:2884 weakly suppresses
- L27:2900 weakly suppresses
- L27:928 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
- 都要
- 都会
- 都能
- 都需要
- 都会有
- 都能够
- 都有着
- 都说
- 都非常
- 都不会
Suppresses
- was
- is
- was
- wasn
- isn
- 的是
- $is
- →is
- (is
- "is