line break splitting a phrase
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. 9 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 line break that cuts a sentence or name in the middle, where the words after the break finish the same thought. Steering did not reveal a clear theme, so the link to the output is unclear.
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.
-
e⏎Southern black tit male, Thornybush Game Reserve,⏎South Africa. [photo Trevor Hardaker ©]⏎Southern black tit female, Thornybush Game⏎Reserve, South Africa. [photo Trevor Hardaker ©]⏎Distribution and habitat
-
⏎the seed at the appropriate time and rate⏎for your species.⏎to spread seed⏎- Measure the area to be planted and determine the total amount of seed⏎-
-
the problem⏎Each member of the team takes a card from the pile in turn (one⏎card each) without the others seeing it.⏎If you have an⏎observer, they might find the following list of things to look for⏎helpful
-
from 'The Photographic Society'⏎in 1892 because of it's lack of interest in art photography. The first show of The Linked Ring Brotherhood⏎in November 1893 was called the Photographic Salon⏎and included
-
DIFFERENCES IN THESE COMMON SLEEP ISSUES⏎Snoring is often one of the first indicators of sleep apnea, but not everyone who snores has⏎What is snoring?⏎Snoring occurs when your throat muscles relax, allowing the
-
and improving the healthcare of⏎Nursing Old Wounds - Recollections⏎by Nancy Leftenant-Colon, as the first black nurse in the Regular Nurse⏎Corps in 1942.⏎The Crile Archives Current Project -⏎
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 key is to focus on the first 10. I think the problem is that the first 10 should focus on the first 10, but the problem is the first 10, the problem is the first
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
extra step is required to terminate the for loop,⏎> hence n + 1 and not n executions), which will consume T4( n + 1 )⏎> time." (from http://en.wikipedia.org/wiki/Analysis_of_algorithms)⏎
-
the Patent Counsel, Johnson Space Center, (281) 483-0837. Refer to MSC-23
-
<|endoftext|>Let’s Talk Fatty Liver Disease⏎When we think about essential bodily organs, there are two primary candidates: our brain and our heart. However, every single organ in the body has a role in something, which makes every organ essential.
-
15⏎Isothermal Community College⏎P.O. Box 804, Spindale NC 28160⏎(828) 395-1307 fax (828)
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
- L8:982 suppresses
- L7:223 excites
- L7:2925 suppresses
- L8:1468 excites
- L8:926 excites
- L6:1540 excites
- L6:181 excites periods in abbreviations and the 's in names
- L8:1396 excites
- L6:199 weakly excites
- L7:355 weakly suppresses
- L7:173 weakly excites
- L7:948 weakly excites
- L6:43 weakly suppresses
- L6:154 weakly excites
- L8:419 weakly suppresses
- L8:1208 weakly suppresses
- L3:439 weakly suppresses
- L7:1354 weakly suppresses
- L4:212 weakly excites
- L7:1771 weakly suppresses
- L8:797 weakly excites
- L8:3048 weakly suppresses
- L7:616 weakly excites
- L5:1962 weakly excites
Feeds later layers
- L12:792 excites
- L13:1108 excites
- L16:2042 excites
- L15:2449 excites line breaks mid-sentence
- L21:348 excites mid-sentence line breaks
- L15:638 excites "the" inside institution names
- L12:1112 excites
- L14:493 excites
- L17:310 excites
- L10:152 excites
- L18:617 excites
- L15:3057 excites
- L10:404 excites
- L12:1823 excites
- L15:646 excites
- L18:933 excites
- L15:2046 excites
- L15:168 excites
- L20:526 excites
- L16:2076 excites
- L15:1846 excites
- L15:2021 excites
- L20:90 excites
- L15:998 excites period inside an abbreviation
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
- ún
- ạch
- EITHER
- чь
- bölgesinde
- ecessarily
- AdminController
- ước
- ChangedEventArgs
- Ẹ
Suppresses
- ⏎
- ctic
- iga
- iae
- iative
- _
- <p
- 徒
- iosa