line break inside a sentence
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. 6 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 hard line break that splits a sentence in the middle, as happens in text copied from older printed or wrapped sources. It is negative on words like "temperatures" after "record high." Turning it up gave unrelated text, so the steering doesn't confirm this.
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.
-
. - I was there too, but didn't⏎see you. I also will be working there from 3 to 5 tomorrow⏎7. When she⏎returns from Moscow, she'll tell us
-
uffed at the knees, a style that⏎involves folding the garment in half and tucking it under the waist. This style⏎of draping the⏎dhoti is known as the half-fold. The dhoti can be
-
all. But finally,⏎after years of waiting, and generally giving up I have run across a passage⏎which may give us the answer.⏎"The formality of⏎the marriage consists in drinking sake after a particular manner. The sake⏎is poured
-
too⏎You can have mutable values in the IO or (ST s) monads, if you need them.⏎> That is, I can't change a⏎> component of somethingNew without creating y. The point is there's⏎>
-
any of the other auberges were situated on a piazza, is not⏎certain but shall be discussed later on.⏎Unfortunately, only three of⏎Cassar's seven auberges have survived. The German aub
-
withdrawn temporarily from Fontaine-les-Croiselles⏎and neutralized fire from Fontaine Wood.⏎No. 2 Company (Ramsay) was⏎with the 6th C.I.B. less the 29th Battalion
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: 11,22,23,24,25,26,27,28,29,30. 30. Sum: 21,22,23,2
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Climate change has an impact on extreme heat waves and record high temperatures, making them more frequent. For example, summer of 2022 brought record high temperatures in Europe. One can compare maximum LST for summer months of 20
-
mph at Austin, TX, and gusts to 75 mph at Tulsa, OK.⏎Many cities in the central and eastern U.S. reported record high temperatures for the date. Highs of 100° at Erie, PA
-
feel tired after use at a computer; vision that is filmy; eyes that are bloodshot; eyes that are irritated or hurt.⏎Recent research has shown that Dry Eye Syndrome (DES) is a chronic inflammatory condition (like arthritis).⏎
-
“Dry Eye” is a term used to describe a group of conditions that cause irritation and damage to the ocular surface. There are many different forms of dry eye, each with a unique cause, and each with a unique set of treatment options
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
- L3:604 excites
- L3:86 excites
- L4:1863 weakly suppresses
- L2:1494 weakly suppresses
- L3:182 weakly suppresses
- L4:863 weakly suppresses
- L4:2241 weakly suppresses
- L5:128 weakly excites
- L3:1563 weakly suppresses
- L3:1198 weakly suppresses
- L1:1464 weakly excites
- L0:1071 weakly excites
- L2:1666 weakly excites
- L5:414 weakly suppresses
- L4:2643 weakly suppresses
- L2:2229 weakly suppresses
- L4:1429 weakly suppresses
- L0:745 weakly excites
- L4:397 weakly excites
- L3:2737 weakly excites
- L2:935 weakly excites
- L3:1558 weakly excites
- L2:1873 weakly excites
- L4:133 weakly excites
Feeds later layers
- L16:151 weakly suppresses
- L17:1242 weakly excites
- L9:2594 weakly suppresses
- L10:1529 weakly suppresses
- L9:2808 weakly suppresses
- L18:863 weakly suppresses
- L16:2337 weakly excites
- L7:133 weakly excites line break inside a sentence
- L15:2995 weakly suppresses
- L11:777 weakly excites
- L7:1728 weakly excites
- L16:735 weakly excites
- L14:2385 weakly suppresses
- L15:446 weakly excites
- L13:1217 weakly excites page and volume numbers in citations
- L16:2437 weakly suppresses
- L24:2236 weakly excites
- L18:1473 weakly suppresses
- L10:3005 weakly suppresses
- L17:1650 weakly excites
- L16:2347 weakly suppresses
- L7:2233 weakly suppresses
- L16:793 weakly suppresses
- L17:2883 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
- >',
- 蘼
- ><![
- heels
- �
- ichi
- igrate
- PermissionsResult
- ämpfe
- реб
Suppresses
- 有效
- 蕉
- hana
- ヽ
- REFER
- 石榴
- 蓬
- GOODMAN
- erase
- ayas