comma before a new clause
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. 21 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 lights up on a comma that comes right after a complete statement, just before the writer starts a new clause such as 'and the second is...' or 'but it...'. When it was turned up, the model fell into long loops of 'and, and, and'.
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.
-
advantages over using a single drive to save data - the first one is redundancy, so if one drive breaks down, the data will be accessible from the remaining ones, and the second is improved performance as the input/output, or reading/writing operations
-
24 and eighth-graders scored 274. The national average for fourth- and eighth-graders was 217 and 263, respectively.⏎Only Connecticut, Massachusetts, Montana and North Dakota scored higher than Vermont’s
-
between the two have been the source of conflict for thousands of years.⏎Both the Jewish and Muslim faith believe in one God. Being monotheistic means just that, the belief in one God. This is the most obvious similarity between the two religions
-
, and Reinforcement Learning.⏎Start with the question, “Do you have a Target Variable?” If the answer is ‘Yes’, it is a supervised learning problem, and if the answer is ‘No’, it
-
AC⏎An airport in Japan will start collecting snow in 2010 to keep the terminals cool during the hot summer months. It sounds a bit crazy, but the small agricultural community of Funagata has been using snow technology since
-
two main blood groups, the ABO group and the Rh group. The ABO group is made up of A, B, AB, or O blood types, and the Rh group is either positive or negative. You blood groups help doctors know
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 current status of the political political political political political political political political political political political political political political political political political political political political political political political political political political political political political political political political political political political political political political political political political political political
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
a series of images of the same blurry face — with the hint that it was a non-American political leader — as it would appear from 200, 100 and 50 meters. At 50 meters,
-
litres of water a day, the UN Food and Agriculture Organisation (FAO) calculates that the water needed to produce a day’s food can range between 2,000 and 5,000 litres.⏎It is hard to
-
⏎Here’s another great example video done by Nick in After Effects using precomps!⏎Check back next week to learn the difference between frame rate and shutter speed, or view more lessons at the Vimeo Video School .<|endoftext|>Some time back I presented
-
00 DPI optical mice as well as 800 DPI laser mice are very popular, but some high-end models are capable of 1000, 1600 or even 2000 DPI tracking speeds.⏎
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
- L19:682 strongly excites
- L20:1358 excites
- L18:366 excites comma inside numbers and dates
- L18:1394 excites
- L20:2082 excites
- L17:116 excites
- L16:240 excites comma after 'sometimes' or 'the truth'
- L14:35 suppresses
- L19:726 suppresses comma inside large numbers
- L18:621 suppresses
- L20:610 suppresses
- L18:174 suppresses
- L17:118 excites comma after a clause ending
- L20:1744 suppresses comma after a sentence's subject
- L19:2451 excites
- L17:1405 suppresses
- L16:1082 suppresses comma linking two clauses after 'and'
- L9:77 excites commas in citations and name lists
- L18:2376 suppresses
- L18:2593 suppresses
- L19:2229 suppresses
- L18:1791 excites
- L16:241 excites
- L20:2730 suppresses
Feeds later layers
- L22:368 strongly excites
- L27:86 strongly suppresses
- L26:22 strongly suppresses the 'res' fragment inside words
- L27:358 strongly excites
- L26:236 strongly excites
- L25:328 strongly excites
- L22:158 strongly suppresses
- L27:130 suppresses
- L27:191 excites
- L27:286 excites
- L25:1221 excites
- L24:32 excites
- L26:80 suppresses
- L26:345 excites
- L25:438 excites
- L26:844 suppresses
- L24:216 excites
- L26:307 suppresses
- L24:199 suppresses
- L26:456 excites
- L27:77 suppresses
- L25:381 suppresses
- L22:1262 suppresses
- L26:1212 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
- ,C
- ,q
- ,default
- ,rp
- ,m
- ,W
Suppresses
- ||⏎
- 善意
- lay
- しました
- "]["
- しておく
- ].[
- emsp
- OwnProperty
- ночь