'to' after verbs of change
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. 16 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 a verb of change or movement and points to where something ends up, like 'warmed up to 3.5 degrees' or 'turn orange to red'. When it was turned up, the model's output got stuck repeating 'to to' and similar 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.
-
racial equality. Indeed, it was the amendment process that created the legal basis to abolish slavery and assure equal political rights for black Americans. As the fight goes on to assert and realize the Constitutional promise of equal representation for all, we humbly
-
flipped and rotated in all directions. This may help to make some furtive words more readable. Also the spacing, boldness and capital case can be switched to make the letters stand out from their context. Optionally a handy hover-cross is shown
-
significant part of Alaska is spared.⏎Permafrost is underlying about 85% of Alaska. Over the past several decades, permafrost has warmed up to 3.5 degrees Fahrenheit and Alaska’s permafrost is now essentially melting
-
izing the aging elementary facilities in its portfolio. The first school being built with net-zero principles is Douglas MacArthur Elementary which first opened in 1943 to serve families who came to the greater
-
2030 that is envisaged in the Sustainable Development Scenario. The buildings sector also remains off track, with emissions rising again in 2019 to an all-time high. Another concerning development was the slowdown in fuel economy improvements around
-
Conditions affecting the tonsils and adenoids⏎Several conditions may affect the tonsils and adenoids in a child:⏎Subscribe to our mailing list to get the updates to your email inbox...<|endoftext|>Detoxification - alcohol; Detox - alcohol⏎
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, from to, and vice a versa, and vice a versa. No, no, no. ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
);⏎void deallocate_model(void *);⏎And that is why I am a header Nazi. Here's how I do<|endoftext|>When you work with people, conflict is bound to happen. The dictionary defines conflict as, “A serious disagreement or argument, typically
-
problem arises when all or most of the benefits of any pollution-emitting activity accrue to one- nation, while all or most of the costs of that activity accrue
-
uler vs. Pie Concept.⏎Webster defines Truth as Reality. But reality to whom? (What does this have to do with consumer research? We will get to that later.) We have often heard the saying that, "Truth is in the
-
988413967-1-768x146-1-300x57.png 0 0 admin https://
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
- L4:328 suppresses 'to' after 'so as' and infinitive cues
- L14:392 suppresses
- L1:2296 excites
- L3:1152 suppresses
- L5:3009 excites 'to' after a verb or noun
- L5:2251 suppresses
- L1:906 suppresses
- L15:78 excites
- L10:43 suppresses "terms" in "in terms of"
- L6:2064 weakly suppresses
- L2:1585 weakly excites
- L6:1534 weakly suppresses
- L4:221 weakly suppresses
- L10:445 weakly excites
- L1:1464 weakly excites
- L4:2047 weakly suppresses
- L0:1551 weakly suppresses
- L10:132 weakly excites
- L6:2543 weakly excites
- L4:862 weakly suppresses 'to' after 'too big' or 'right angles'
- L1:146 weakly excites
- L5:418 weakly suppresses
- L6:211 weakly excites
- L0:178 weakly suppresses
Feeds later layers
- L17:2821 strongly excites
- L17:1297 excites
- L19:1874 excites
- L18:1408 excites
- L18:1001 excites
- L23:373 excites 'to' after a noun or adjective
- L19:858 excites
- L17:2527 excites
- L21:2012 excites "to" before an infinitive or noun
- L18:2760 excites 'to' after a verb of sharing or using
- L17:1843 excites
- L25:286 excites
- L17:472 excites
- L21:1657 excites
- L18:285 suppresses
- L21:1081 excites
- L18:90 excites
- L17:303 excites
- L19:1237 suppresses
- L17:1884 suppresses
- L18:1675 weakly excites
- L18:176 weakly excites
- L19:1463 weakly suppresses
- L17:1122 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
- /from
- 的就是
- (screen
- gether
- yal
- ulet
- render
- ombat
- validate
- izi
Suppresses
- ARED
- heit
- IFIED
- lation
- ified
- lijk
- roleum
- 扫一扫
- ification
- -fat