unclear: list items and sentence breaks
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 mixed: they peak on different tokens. 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 on a mix of things: list dashes, the end of instruction sentences, and some technical headings. Turning it up made the model output 'En' and 'El' fragments, which does not clearly connect to those examples.
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.
-
Rake Angle⏎- Positive and Negative Rake Angle in End Milling⏎- Relationship Between Cutting Tool Angle, Rake Angle, and Relief Angle⏎- Problems in Conventional Rake Angle Measurement⏎- Solution to Problems in Rake Angle
-
water, according to the instructions on the label⏎- Use a sponge or rag to apply the detergent solution to the mouldy surface. Scrub the area⏎- Thoroughly dry the surface by<|endoftext|>Home > Technical Articles⏎A person new
-
1970 UNESCO Convention on the Means of Prohibiting and Preventing the Illicit Import, Export and Transfer of Ownership of Cultural Property was ratified by the
-
according to the U.S. Drug Enforcement Administration (DEA).⏎Ketamine Found Effective, but More Study Suggested⏎The researchers found that ketamine significantly reduced pain scores in 3
-
, bladder infections, pneumonia, gonorrhea, and E. coli or salmonella infection.⏎How to use⏎Use Ampicillin as directed by your doctor.⏎Take Ampicillin by mouth at least one-half hour before or 2 hours
-
Cover seed to a depth of 1/16 to 1/8 inch by raking it in and lightly⏎rolling or firming the soil.⏎- The finished seedbed should have shallow uniform depressions about⏎1/2 inch
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: what if we reposition the research as a form of research for research? research research research research research research research research research research research research research research research research research research research research research research research research research research research research research research research research research
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
there is no proof for this. https://www.academia.edu/39903340/Displaying_an_Icon_The_Mosaic_Icon_of_Saint_Demetrios_at_Sassoferrato_and_
-
New England Emigrant Aid Company⏎Date: June 9, 1857⏎Charles H. Branscomb, acting on behalf of the New England Emigrant Aid Company, contracted with Abner Doane to construct a school house
-
Usage dialog remains open), or|⏎|·||Click on the details for a diagram and click on the Open button to open that diagram and highlight the element (the Element Usage dialog closes)|⏎The material ("material" includes all pages, documents
-
(t)=Vmax×sin(ωt+ΦV) and current I(t)=Imax×sin(ωt+ΦI) where Vmax, Imax are the peak values, ω is the common angular pulsation and
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
- L19:484 strongly excites
- L23:278 strongly excites
- L19:474 strongly excites
- L25:198 strongly suppresses
- L21:185 strongly suppresses
- L19:1941 strongly excites
- L19:1485 strongly suppresses
- L20:351 strongly excites
- L20:473 strongly suppresses
- L19:2730 strongly suppresses
- L21:335 strongly suppresses
- L19:1455 strongly suppresses
- L21:846 strongly excites
- L18:2824 suppresses
- L19:96 excites
- L25:594 excites
- L20:308 excites
- L21:370 excites
- L24:512 suppresses 'the' inside organisation and convention names
- L24:192 suppresses
- L19:68 excites full stops closing a claim
- L20:1002 suppresses
- L20:296 suppresses
- L22:1829 excites
Feeds later layers
- L27:2632 suppresses
- L27:387 suppresses 'a' right after 'played' or 'play'
- L27:455 excites
- L27:53 suppresses
- L27:1745 excites
- L27:1507 excites
- L27:137 weakly excites
- L27:715 weakly excites
- L27:2097 weakly excites
- L27:787 weakly excites
- L27:16 weakly excites
- L27:49 weakly suppresses
- L27:601 weakly excites
- L27:1599 weakly excites
- L27:12 weakly excites
- L27:918 weakly excites
- L27:229 weakly excites
- L27:1230 weakly excites
- L27:852 weakly excites
- L27:133 weakly excites
- L27:822 weakly excites
- L27:316 weakly excites
- L27:178 weakly suppresses
- L27:1361 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
- En
- El
- Es
- El
- O
- Ex
- Em
- O
- Ed
- Input
Suppresses
- atory
- ämp
- asting
- aturing
- utory
- ucky
- ทน
- fähig
- سبة
- ないので