comma after a speech verb
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. 2 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 fires on a comma that comes right before a quotation, such as "He says," or "You may be asking yourself,". When it was turned up, the writing did not clearly show this habit.
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.
-
(OLST)1,2 is a simple, easy and effective method to screen for balance impairments in the older adult population.⏎You may be asking yourself, "how can standing on one leg provide you with any information about balance, after
-
.⏎What’s the Best Humidity Level for your Home? How do you maintain the level throughout the year?⏎You may have thought, “What is the ideal level for my house in the summer or winter months?”
-
watches Gertrude's reaction to the statement. By her reaction, he may be able to judge whether she has knowledge of Claudius' crime. He says, "Almost as bad...As kill a king, and marry with his brother."
-
isticexpialidocious?"⏎- Our speaker, sensing our sensory overload, steers us back on course. At the end of line 9, he says, "the gaps I mean," reminding us of the matter at hand (the broken
-
).⏎It is interesting to note, however, that while the Torah commands us to verbally recall the Exodus from Egypt twice
-
to gauge.⏎Recent technological and scientific advances, particularly the recovery and sequencing of ancient DNA from old bones, are revealing some of that lost history.⏎We now suspect, for example, that the bacterium has been infecting people since at least the
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: I have been working on the idea of a "data" and "data" is the same as a "data" and "data" is the same as a "data" and "data" is the same as a "data"
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
indirect techniques only calculate it from other indirect indicators.⏎Image Credit: Soonthorn Wongsaita / Shutterstock⏎Many devices for LOD measurement use the thermogravimetric principle. This uses the total loss of weight incurred by the sample on
-
He professes to know how teachers should teach in the entire country and what they should teach, and yet, he is not a teacher. Coleman worked for McGraw-Hill (textbook publishing company) before leaving to create GROW a curriculum
-
of the cathedral in the Romanesque or Norman style, starting with the nave. Robert died shortly afterward, but the work was carried forward by Geoffrey de Montbray, his successor as bishop, appointed in 1048. Geoffrey
-
at bottom left showing two islands discovered by Cook and Marion du Fresne. One of the few maps in the volume drawn by Bonne's son. Engraved by Andre. Condition: A nice impression with minor offsetting and a few faint
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
- L0:714 weakly excites
- L0:696 weakly suppresses
- L1:126 weakly excites
- L0:729 weakly suppresses 'st' in dates like 'June 1st'
- L1:2796 weakly excites
- L1:2342 weakly excites
- L1:1125 weakly suppresses
- L1:1467 weakly suppresses
- L0:2968 weakly suppresses
- L1:2570 weakly suppresses
- L0:251 weakly suppresses
- L1:2736 weakly excites
- L0:173 weakly suppresses
- L1:915 weakly suppresses
- L0:263 weakly suppresses
- L1:909 weakly excites
- L0:1472 weakly excites
- L0:1524 weakly suppresses
- L1:2119 weakly excites
- L0:2204 weakly suppresses
- L0:2774 weakly excites
- L1:384 weakly suppresses
- L1:445 weakly excites document start marker token
- L1:507 weakly excites
Feeds later layers
- L7:1833 excites
- L6:700 weakly excites
- L6:600 weakly excites
- L11:2311 weakly excites
- L4:622 weakly suppresses
- L4:2378 weakly excites
- L16:3066 weakly excites
- L18:647 weakly suppresses
- L8:2936 weakly excites
- L3:1537 weakly suppresses
- L8:1105 weakly excites
- L4:2605 weakly suppresses
- L25:860 weakly excites
- L11:2007 weakly suppresses
- L10:2026 weakly excites
- L8:1312 weakly excites
- L20:2412 weakly suppresses
- L4:679 weakly excites
- L7:1081 weakly excites
- L8:2957 weakly excites
- L15:1745 weakly suppresses
- L13:1145 weakly excites
- L13:1302 weakly suppresses
- L16:2796 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
- enes
- leen
- 切尔
- adamente
- enet
- ewise
- 人心
- лим
- GenericType
- contracting
Suppresses
- avery
- łów
- trạng
- ớt
- hog
- itution
- itive
- berg
- 是一
- ición