"to" before an infinitive or noun
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. 21 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 comes right after a verb or noun and leads into the next part of a sentence, like "what we believe to be" or "put to death". Turning it up did not produce a clear theme; the text just drifted.
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.
-
the attributes of an educated person. The UMass Dartmouth Commitment to Student Learning, which is the foundation of the General Education Program, captures what we believe to be the essence of the UMD educated person.⏎UMass Dartmouth Commitment
-
"The median response indicated a 23 percent increase in productivity in the future atmosphere," said ORNL's Rich Norby, lead author of the paper to be published Dec. 13 in the Proceedings of the National Academy of Sciences
-
by a method unchanged since the age of the Moorish invaders _ swamping them under a flood of water from the local canal.⏎But those days may be coming to an end. Martorell is one of a growing number of Spanish farmers who
-
preventable cause of harm. To improve knowledge on anticoagulant safety, the United States (US) Office of Disease Prevention and Health Promotion issued a call to generate more evidence on anticoagulant drug interactions.⏎Potential drug-drug interactions
-
domino reaction across connected devices. In no time at all, data infrastructures can be destroyed, making the IoT one of the biggest data vulnerabilities in the years to come.⏎2. Mobile malware⏎While malicious apps
-
13th century Christian Europe willingly looked upon him as an international judge.⏎He was renowned for his charity. “The peace and blessings of the realm come to us through the poor,” he would say. Beggars were fed from his
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: When the match get get meet at a match, when the match meet match meet meet meet meet meet match meet meet meet meet meet match meet meet meet meet match meet meet meet meet match meet meet meet meet match meet meet meet meet meet match
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
.S. history. The attack on the Greenwood section of Tulsa, home to a prosperous African American community in the segregated city, left over 35 blocks burned to the ground and over 10,000 people — most of them
-
their cows to be fed. In the back of everyone’s mind is the fact that the common is going to be ruined because the grass is going to be eaten to depletion. Still, everyone wants to get grass for their cows. No one thinks
-
space will likely always grapple with.⏎In May 2020, Rhys’ childhood home in Dominica's capital, Roseau, was demolished to accommodate commercial construction, an act that Ramlochan found “instructive”
-
towns of Hardwick and Petersham before becoming its own separate town in 1801. The town was abandoned and leveled in 1938 to make way for the Quabbin Reservoir. Much of the town, including
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
- L18:1001 excites
- L18:2760 excites 'to' after a verb of sharing or using
- L17:1297 excites
- L19:858 suppresses
- L6:2000 suppresses
- L9:1890 excites
- L16:1168 excites 'to' after verbs of change
- L4:1035 suppresses 'for' after verbs like 'fall' or 'click'
- L19:1769 excites
- L19:2156 suppresses
- L20:810 suppresses
- L20:2814 suppresses
- L19:2535 excites
- L12:2212 weakly suppresses 'to' or 'for' opening a purpose phrase
- L20:2262 weakly excites
- L2:24 weakly suppresses
- L0:318 weakly excites
- L7:1584 weakly suppresses
- L0:2193 weakly excites
- L20:1978 weakly suppresses
- L5:2251 weakly suppresses
- L20:1623 weakly excites
- L16:1939 weakly excites line break mid-sentence in wrapped text
- L1:1900 weakly excites
Feeds later layers
- L22:1166 suppresses 'to' after words needing an infinitive
- L24:2854 suppresses
- L22:1806 suppresses "to" introducing a purpose clause
- L24:2768 suppresses
- L22:266 excites
- L22:627 suppresses
- L23:1169 suppresses
- L22:228 weakly excites
- L23:1115 weakly suppresses
- L22:200 weakly suppresses
- L23:1378 weakly suppresses
- L22:2869 weakly suppresses
- L24:265 weakly suppresses
- L24:1072 weakly suppresses
- L23:566 weakly suppresses
- L25:364 weakly suppresses
- L26:296 weakly suppresses
- L24:442 weakly suppresses
- L23:887 weakly suppresses
- L26:189 weakly excites
- L22:603 weakly suppresses
- L24:1325 weakly suppresses
- L25:1197 weakly suppresses
- L24:208 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
- Swipe
- 为
- 为进一步
- 是如何
- with
- cribes
- 为目的
- 起来了
- whose
- stdbool
Suppresses
- 避免
- avoid
- 防止
- 适应
- 期
- 达到
- avoid
- 逃避
- avoided
- achieve