middle pieces of split names
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. 18 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 a fragment in the middle of a longer name or word that has been chopped into pieces, such as the 'k' in a surname or the 'ev' in a Russian name. Turning it up made the text ramble without a clear theme.
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.
-
the ridge was a narrower valley, with the Fedhoukine hills to the north.⏎It was towards the latter part of that memorable day
-
ing some men honor, glory and fame,⏎but others only shame and hardship.⏎This I can say for myself:⏎that for awhile I was the Heodeninga<|endoftext|>This essay has been submitted by a student. This is not an example of
-
entire system, each part at a different lab is connected through the Internet. Then we can understand how the earthquake impacts entire buildings.”⏎According to Joy Pauschke, an NSF official who helped plan the first experiment, the idea for NEES
-
East Timor, Ethiopia, and Eritrea Chapter 19 What if Patriarchy is The Big Picture? An Afterword⏎PortraitDyan Mazurana is senior research fellow at the Feinstein International Famine Center, Tufts University.
-
, Fluxys, Gasunie, GRTgaz, Open Grid Europe, Snam and TIGF
-
Survived first manned spaceflight abort during launch. More...⏎Filipchenko Filipchenko, Anatoli Vassilyevich (1928-) Russian pilot cosmonaut. Flew on
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: what if it is possible to create a number in a number line with the same number of digits in a number that is not exactly 3, which is a number that is less than 1000000000
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Select answer A for those questions.)⏎What’s that? You don’t understand what we mean by selecting answer A? It’s OK. You’re going to P-R-A-C-T-I-C-E and get very comfortable with the test format so you
-
fish hold most all of the current world records from 12- to 130-pound test line for both men and women. The all tackle I.G.F.A. record is 494 pounds on
-
. As you write her words, bring her attention to the use of uppercase and lowercase letters and spacing between words.⏎“C-a-r car. That’s the word car. Now I’m going to leave a space
-
definitely attend? There’s the Wild World of Animals that features 14 interesting animals—from scorpions to legless lizards—sponsored by N.A.T.U.R.E., Pace’s Environmental Club. Interested in photography? There will be
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
- L17:324 strongly excites
- L14:33 strongly suppresses
- L16:405 strongly suppresses
- L15:661 strongly excites
- L9:39 suppresses
- L13:695 excites
- L15:94 suppresses
- L15:442 suppresses
- L16:1138 excites
- L16:1432 suppresses
- L6:182 suppresses
- L16:1319 suppresses
- L15:1690 suppresses
- L14:921 excites
- L1:344 suppresses
- L16:2141 excites
- L16:337 excites
- L6:248 excites
- L17:1285 excites
- L15:486 excites
- L15:1137 excites
- L16:1064 suppresses
- L15:984 excites
- L16:363 excites
Feeds later layers
- L19:352 strongly excites
- L20:214 suppresses
- L21:122 excites
- L20:705 excites
- L20:494 excites
- L19:904 suppresses
- L21:39 suppresses
- L19:224 excites
- L19:667 excites
- L21:36 suppresses
- L20:794 excites
- L20:2723 excites
- L19:2764 excites
- L21:129 excites
- L20:1108 excites
- L19:402 suppresses
- L19:651 excites
- L19:75 suppresses
- L19:737 weakly suppresses
- L20:154 weakly suppresses
- L19:2902 weakly suppresses
- L21:13 weakly suppresses
- L19:898 weakly suppresses
- L20:1504 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
- cient
- unaware
- chin
- bones
- catch
- ์
- ulo
- Ign
- Process
- batim
Suppresses
- INGTON
- [Unit
- ington
- [unit
- ernaut
- olini
- stress
- face
- enthal
- rypted