unclear: end-of-text markers and word fragments
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 fires most strongly on the hidden marker that signals the end of one document, right where an unrelated text would begin. It is hard to pin down beyond that. When it was turned up, the model's replies drifted into unrelated topics, with no clear theme.
Written by Claude from this neuron's activations and steering outputs. Confidence: low. 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.
-
Siberian ginseng, improves stamina and enhances immune function, as do reishi mushrooms — another adaptogenic remedy. Ashwagandha, an Indian<|endoftext|>Moral FriendSeries teasers are where you try to complete the sequence of a
-
data. Further, the cloud and artificial intelligence have a symbiotic relationship. The cloud handles the fundamental source of information that augments artificial intelligence networks, and artificial<|endoftext|>If I were to think of an emblematic phrase, in arts education, it
-
4, 1957, Sputnik, the Soviet satellite that was the first to orbit Earth, shocked Americans and delighted the Soviets. A month<|endoftext|>O’Connell, Caitlin and Donna M. Jackson. The Elephant Scientist.
-
, and encouraged trade and industry, dispatching a seaborne trading expedition to Punt, a land on the Somali coast of Africa, and exploiting the copper<|endoftext|>Industrial design is a process of design applied to products and industrial goods that are to
-
Prev Med Rep⏎Authors: Altman EA, Lee KL, Hecht CA, Hampton KE, Moreno G
-
queen Violant sought support for her grandchildren from her brother, King Peter III of Aragon, who agreed to protect and guard them in the kingdom of Aragon<|endoftext|>ALR provided recommendations that helped inform the national Childhood Obesity Task Force action plan,
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: In the context of the problem, the user is asking for the sum of the digits of the product, so we need to compute 10^k, for k=1, 2,..., 1000. But
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
blackberries, pears, raspberries, strawberries, whole grain breads and bran cere
-
gardenia.⏎Insecticides kill off the whitefly’s natural predators. These predators include
-
side effects”⏎While the changes are only a few percent, with more than 40
-
Porcelain⏎Process – Raw Porcelain is rolled to create the tile, each letter is
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
- L4:1411 strongly excites
- L8:269 strongly excites
- L3:1392 strongly excites
- L8:624 strongly excites
- L7:1130 strongly excites
- L2:1519 strongly suppresses
- L7:406 suppresses
- L3:921 suppresses
- L2:182 suppresses
- L8:313 suppresses
- L6:927 suppresses
- L7:234 excites
- L7:241 suppresses
- L5:112 suppresses
- L5:1341 excites
- L7:1866 excites
- L8:1053 suppresses
- L2:2010 excites
- L8:154 excites
- L2:699 excites
- L6:2834 excites
- L7:911 suppresses
- L8:49 weakly excites
- L1:1337 weakly suppresses
Feeds later layers
- L12:978 strongly excites
- L10:710 strongly excites
- L10:369 strongly suppresses
- L11:1143 strongly suppresses
- L12:354 strongly suppresses
- L10:762 strongly suppresses
- L12:933 suppresses mid-sentence 'also' and 'another'
- L12:2162 suppresses
- L12:2508 suppresses
- L14:3021 suppresses
- L12:2618 suppresses document boundary token
- L15:3039 suppresses
- L12:1302 suppresses
- L14:2862 suppresses
- L15:2008 excites
- L17:2032 suppresses
- L10:1749 excites
- L16:2255 suppresses
- L17:658 excites
- L14:1023 suppresses
- L11:1719 suppresses dash starting a bullet item
- L10:286 excites
- L15:2742 suppresses
- L20:2391 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
- thon
- 名叫
- �
- 以下是
- 金银
- ...⏎⏎⏎⏎
- 牌子
- (_,
- 到底
- JECT
Suppresses
- herself
- онт
- auer
- ост
- rame
- himself
- Himself
- orrent
- át
- ienda