mixed: 'New' in place names, square roots
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. 24 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 mixed. It mainly reacts to the word 'New' in place names like New York or New Mexico, and also to square-root signs in maths. Turning it up or down produced repetitive text about 'has' and 'have', which matches the words it tends to suppress.
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.
-
limit it gives you the function being integrated.⏎Thus (because the lower limit doesn't play any role being a constant number)⏎dF(x)/dx2 = √(1 + t3) where t = x2 (2)⏎If
-
the speeds of two satellites placed in stable circular orbits of radius R and 2R respectively?⏎C) 1 / 2⏎D) 1 / √2⏎E) 2 √2⏎What is the centrip
-
lenker, a Senior Researcher at Institut Jean-Nicod within France's National Center for Scientific Research (CNRS) and a Global Distinguished Professor at New York University.⏎The combined team of linguists and primatologists analyzed alarm calls
-
Biology at the University of New Mexico. Troy L. Best is a professor of biological sciences at Auburn University.
-
tutorial from the Australia Telescope National Facility, Parkes, New South Wales⏎Amateurfunk, Erde-Mond-Erde,
-
detection being between 60 and 70 years.⏎“Cancer is a disease of the aging,” said Dr. Deborah Axelrod, associate professor at New York University (NYU) Langone Medical Center’s Division of Breast Surgery.
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: having a has a has a has a has a has a has a has a has a has a has a has a has a has a has a has a has a has a has a has a has a has a has a has a
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Edison (1847-1931) in 1878. Edison redesigned the telephone shortly after its invention by Alexander Graham Bell (1847-1922) in 1876.
-
<|endoftext|>West Virginia Population 2017⏎Located in the central eastern sector of the United States of America, West Virginia was admitted into the Union in 1863. It is the 41st largest and the 3
-
and wear appropriate clothing.⏎Suspects can invoke the Sixth Amendment right to counsel during lineups and showups held after they have been arraigned or indicted. They have the right to have an attorney present, but the attorney only participates as an observer
-
and quitting any form of tobacco use are ways to prevent or delay the onset of type 2 diabetes. In a healthier world, everyone, regardless of socioeconomic status, would have access to sustainable, affordable and healthier diets
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
- L21:162 excites
- L21:2804 excites
- L22:2524 excites
- L21:2212 suppresses
- L22:2752 suppresses
- L20:2854 suppresses 'official' and government institution words
- L19:2900 excites
- L14:1884 excites
- L15:559 excites
- L19:1959 weakly suppresses
- L17:2930 weakly excites
- L21:2448 weakly suppresses "a" after give or get verbs
- L22:1402 weakly suppresses
- L22:2775 weakly excites
- L16:1367 weakly excites
- L19:1371 weakly suppresses
- L18:2198 weakly suppresses
- L12:1136 weakly suppresses
- L22:3060 weakly excites
- L19:377 weakly suppresses
- L11:374 weakly excites
- L23:387 weakly suppresses
- L9:2950 weakly suppresses
- L19:1461 weakly suppresses
Feeds later layers
- L27:756 strongly suppresses
- L26:2817 suppresses
- L25:613 excites
- L25:685 suppresses
- L25:608 weakly excites
- L27:1095 weakly suppresses
- L27:564 weakly excites
- L26:980 weakly excites
- L25:929 weakly suppresses
- L25:1671 weakly excites
- L26:2563 weakly suppresses
- L26:2875 weakly excites
- L25:1873 weakly suppresses
- L25:2702 weakly excites
- L27:2364 weakly excites
- L25:135 weakly excites
- L27:419 weakly excites
- L27:2469 weakly excites
- L25:1623 weakly excites
- L26:2349 weakly excites
- L26:1914 weakly suppresses
- L26:2462 weakly excites
- L26:648 weakly suppresses
- L25:598 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
- inely
- itti
- orsi
- idia
- ibili
- nah
- станавли
- inaire
- outines
- mitted
Suppresses
- 有
- has
- have
- has
- (has
- .has
- having
- Have
- had
- have