capital N starting an abbreviation
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. 25 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 responds to a capital letter 'N' (or 'H') that starts an abbreviation or name, such as NDR, NCI or N.Y.U. Its opposite side reacts to the 'im' in Cimarron. Turning it down made the model repeat the word 'with' endlessly, and turning it up gave unrelated text.
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.
-
study from China, but hope that researchers are on the right track and a solution is in sight gauge, as a virologist Christian Drosten of the NDR Podcast "The Coronavirus-Update" reveals.⏎In the study from Beijing,
-
, whilst on its way to meet me in London.⏎Finland is essentially a flat country, slightly mountainous towards the north, but even her highest peak (Haldesjock, in Finnish Lapland) is under 4,0
-
ism. Paton also illustrates this theme through the development of several characters in the novel: the literal move of characters such as the pregnant girl to rural life in Ndotsheni represents a change to a greater moral sense, while
-
enough clinical verification and validation to know and feel comfortable that what we’re detecting with them is clinically meaningful,” said Lynn Sorbara, Ph.D., of NCI’s Division of Cancer Prevention.⏎Different Tests for Different Tumor Molecules⏎
-
to monitor the research on BPA. Glass bottles, however, also leach chemicals, including aluminum, explains a study by Clemens Reimann of the Norges Geologiske Undersøkelse in Norway and colleagues. Aluminum is
-
however, was strangely reassuring.⏎“Anxiety disorders are the most common condition in psychiatry,” said Dr. Naomi Simon, professor of psychiatry at N.Y.U. School of Medicine and director of the Anxiety and Complicated Grief
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: with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Digital data sets that describe aquifer characteristics of the alluvial and terrace deposits along the Cimarron River from Freedom to Guthrie in northwestern Oklahoma⏎U.S.
-
ette contains digitized aquifer boundaries, maps of hydraulic conductivity, recharge, and ground-water level elevation contours for the alluvial and terrace deposits along the Cimarron River from Freedom to Guthrie in northwestern Oklahoma. Ground water in
-
overlie the Permian-age Formations.⏎The aquifer boundaries are from a ground-water modeling report on the alluvial and terrace aquifer along the Cimarron River from Freedom to Guthrie in northwestern Oklahoma and published digital surf
-
After a seismic swarm, discussed below , the initial strombolian activity localizes from April 2 on a crack on the southwest flank of Pico do Fogo . A cinder cone called Pico Pequ
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
- L20:2338 suppresses
- L20:3030 suppresses
- L19:866 suppresses
- L21:665 excites
- L19:1437 excites
- L23:839 suppresses comma before a trailing 'with' clause
- L23:1437 weakly excites
- L22:2101 weakly excites
- L21:1597 weakly excites
- L14:1323 weakly suppresses
- L19:1036 weakly suppresses
- L12:2994 weakly suppresses
- L22:2157 weakly suppresses
- L18:2811 weakly suppresses
- L10:1481 weakly excites
- L17:1044 weakly suppresses
- L1:1236 weakly excites
- L18:2321 weakly suppresses
- L17:1659 weakly suppresses 'as' after 'use' or 'taking' (as a model)
- L18:251 weakly suppresses
- L19:2292 weakly suppresses
- L8:1450 weakly suppresses
- L2:1234 weakly excites
- L14:450 weakly excites
Feeds later layers
- L26:741 strongly suppresses
- L27:603 strongly suppresses
- L27:1443 suppresses
- L27:2153 suppresses
- L26:223 excites
- L26:973 weakly suppresses
- L26:82 weakly suppresses
- L26:309 weakly suppresses
- L26:2350 weakly suppresses
- L27:102 weakly excites
- L27:1069 weakly excites
- L26:1293 weakly excites
- L26:719 weakly suppresses
- L26:2058 weakly excites
- L27:491 weakly excites
- L26:2764 weakly excites
- L26:1 weakly excites
- L27:1872 weakly excites
- L26:1640 weakly excites
- L26:1124 weakly excites
- L26:1836 weakly excites
- L27:74 weakly suppresses
- L26:1902 weakly excites
- L27:277 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
- rocessing
- agas
- edes
- otyping
- ETING
- ONSE
- yms
- istration
- eding
- utomation
Suppresses
- with
- with
- With
- With
- _with
- →with
- WITH
- avec
- WITH
- -with