unclear: document ends and math commas
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. 23 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 is hard to pin down. It fires most strongly where a web page's text seems to cut off and the next document begins, and it goes strongly negative on commas in maths problems. Turning it down made the model repeat the word 'find' and turning it up gave ordinary sentences, so the pattern is not clear.
Written by Claude from this neuron's activations and steering outputs. Confidence: low. 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.
-
31 minutes outside the lab.⏎After what was then a record 84 days in space, Carr, Pogue and Gibson bid farewell to Skylab and<|endoftext|>By
-
joined to form a⏎quadrilateral. What do you notice about this quadrilateral as the⏎dynamic image changes? Can you prove your conjecture?⏎It is impossible to trisect an angle using
-
decision.⏎Each branch must depend on the other in order to keep things running smoothly. Many times, this presents challenges along the way. It also means that it<|endoftext|>Writers Building Kolkata⏎Writer’s Building, Kolkata⏎Going Back to the History
-
so that cows can access shade and water whenever they want it. But that is easier said than done. Over the next ten years, we will be expected to<|endoftext|>
-
. Their music and lyrics are synonymous with the 60’s. From “The Sound of Silence” to “Bridge Over Troubled Water”, their songs act<|endoftext|>Read Music Notes⏎How to Read All the Music Notes -- With Ease⏎So
-
resistant” may be a special treat to hungry deer, as they munch on tender soft new growth. Like many other animals, deer have food they avoid, and<|endoftext|>Seven miles to the south west of Sheffield there's a piece of wild moor
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: find find find find find find find find find find find find find find find find find find find find find find find find find find find find find find find find find find find find find find find find find find find find find find find find
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
all variables and given/known data If an athlete jumped 2 feet high and left the ground at an angle of 20 degrees with respect to the horizontal, how fast was the athlete going in the forward (positive horizontal) and upward (
-
centripetal acceleration at point B?⏎Fig.1 above refer to a point moving along a circular path. If BD is the diameter of the circle of rotation, then the velocities at A and B are
-
.⏎A straight line is drawn between the end points of the arc would be a⏎chord of the circle.⏎If the arc length is exactly half the circle, this called a semicircular arc. See Semicircle definition.⏎Naming and
-
.⏎An ideal monatomic gas initially at 298 K and 1 atm pressure is⏎expanded from a volume of 1 liter to 2 liters. Consider separately⏎isobaric, isothermal and adiabatic expansions to
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
- L17:2296 suppresses
- L18:1586 suppresses
- L15:1416 suppresses
- L21:728 suppresses
- L19:877 excites
- L20:2335 excites comma ending an 'if' clause
- L3:66 suppresses
- L7:61 excites
- L18:2010 excites
- L3:52 suppresses
- L9:128 suppresses
- L19:2676 excites
- L20:810 suppresses
- L15:2327 suppresses
- L11:1483 suppresses
- L4:150 weakly suppresses
- L13:2989 weakly excites nouns for who or what a text is about
- L6:431 weakly suppresses
- L9:156 weakly excites
- L8:584 weakly excites
- L8:49 weakly suppresses
- L11:1843 weakly excites
- L20:1086 weakly suppresses
- L12:1069 weakly excites
Feeds later layers
- L24:1302 suppresses
- L27:1284 suppresses
- L26:1282 suppresses
- L27:247 suppresses
- L25:361 suppresses
- L26:623 suppresses
- L25:364 weakly suppresses
- L26:1422 weakly suppresses
- L26:675 weakly suppresses
- L24:837 weakly suppresses
- L26:2988 weakly suppresses
- L24:290 weakly suppresses 'to' in set phrases like 'to come'
- L25:1335 weakly suppresses
- L26:1104 weakly suppresses
- L24:2640 weakly suppresses
- L26:939 weakly suppresses
- L26:671 weakly suppresses
- L27:900 weakly suppresses
- L27:645 weakly suppresses
- L26:296 weakly suppresses
- L26:2821 weakly suppresses
- L24:2579 weakly suppresses
- L25:660 weakly suppresses
- L24:2716 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
- yst
- ivalence
- 本报讯
- aves
- ibox
- ipped
- 批
- offee
- 깨
- ierten
Suppresses
- find
- find
- 求
- Find
- Find
- .find
- _find
- (find
- -find
- determine