words introducing an example or aside
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. 8 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 fires on small connecting words that introduce an example or a side comment, such as 'for example', 'for instance', 'like', 'though' and 'or'. Turning it up did not produce a clear change in the model's writing style.
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.
-
we should be including in the language guide.⏎Throughout EmbraceU we have decided to use medical words (like mouth, anus, vulva, vagina, and penis) to describe body parts
-
was to break out a screwdriver and fiddle around with the device. Even then, there were only a few different things that could be changed. Now, though, there are hundreds
-
which is addictive. Public health advocates warn that e-cigarette makers are using the same strategies that tobacco companies employed to attract young people — candy flavorings, for example. Lax state-level regulations also make it easy for young people to buy them
-
inflation rate over the last seven months. This exercise reveals whether the increase in inflation since last fall is due primarily to a few outliers—food and energy, for instance—or whether it reflects a broader acceleration in prices.⏎A “Heat-Map”
-
, called added sugars, are viewed as a detriment to a healthy diet. Maximizing sugar’s benefits requires balancing the healthier and less wholesome sources.⏎Sucrose, or table sugar, is the main source of sugar in most American diets. It consists
-
collaborative coding environment can be conveniently captured by a graph. The vertices (or things) are people, organizations, projects, and tickets. The edges (or
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: the number 14. So, the number 14 is in the equation. Now, the problem is to find the value of the equation. I will also consider the time limit of 120 seconds. The equation is
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
berspace – being important part of the implementation of human rights online.⏎One of the most popular types of network communities today is knowledge exchange communities (practice communities) that represent social communities, members of which are involved in collaboration and the crucialcondition here
-
. To January 1, 1944, the Bartlesville sand has produced 7,566,000 barrels of oil from 5,610 productive acres and 871 oil wells.
-
precipitation for the year in Mount Surprise is 31.0" (787.4 mm). The month with the most precipitation on average is January with 8.0" (203.2 mm) of precipitation.
-
was crowned as the first king of Jerusalem on Christmas Day, 1100, by Dagobert of Pisa, the Patriarch, who would have preferred to turn the kingdom into a theocracy. In the spring of
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
- L7:337 excites
- L6:1389 excites
- L3:372 excites
- L7:1222 suppresses
- L3:487 suppresses
- L7:2811 excites
- L2:225 excites
- L7:65 excites
- L3:353 excites
- L2:263 suppresses
- L6:1387 excites
- L7:497 suppresses
- L7:3071 weakly excites
- L7:2503 weakly excites
- L7:1170 weakly suppresses
- L7:84 weakly excites
- L5:1067 weakly suppresses
- L6:327 weakly excites
- L7:506 weakly excites
- L7:658 weakly suppresses
- L0:843 weakly excites
- L6:1006 weakly excites
- L5:974 weakly excites
- L5:1623 weakly suppresses
Feeds later layers
- L9:1370 suppresses
- L10:1985 suppresses
- L10:1459 suppresses
- L9:1113 suppresses
- L10:271 excites
- L14:602 suppresses
- L10:1392 excites
- L12:982 suppresses
- L12:2240 suppresses
- L14:1181 suppresses
- L13:637 suppresses opening parenthesis after a term
- L12:2882 weakly suppresses
- L14:2820 weakly suppresses "and" or "or" linking two counted items
- L13:2071 weakly excites
- L11:1758 weakly suppresses
- L10:894 weakly suppresses
- L14:2964 weakly suppresses
- L9:671 weakly suppresses
- L10:1580 weakly suppresses
- L15:1370 weakly suppresses
- L21:2324 weakly suppresses "in" after "shown" (figure references)
- L9:630 weakly suppresses
- L17:2896 weakly suppresses
- L13:691 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
- raf
- phia
- alink
- ongan
- etus
- olatile
- 这张
- fcc
- vfs
- льн
Suppresses
- 荡
- endum
- ぎ
- 頓
- towards
- endcode
- 보
- �
- upon
- Sho