sentence-ending periods before new sentences
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. 11 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 the full stop or line break that ends a sentence, right before a new one starts, especially after a name or place. When it was turned up, the text drifted into odd, unrelated topics, so it does not clearly steer toward one theme. When turned down, it produced repetitive text.
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.
-
Year 2.⏎Year 1 follow Curriculum Music⏎Year 2 follow ‘Listen 2 Me
-
Naval Cemetery. You can still visit a memorial erected in their honor in Hagatna today.⏎
-
Research Report. Book 4.<|endoftext|>Sardar Sarovar Dam, a project for
-
1944.⏎Others are Kolmaneskuppe, Namibia, a former diamond mining town whose
-
state of Utah. As of the 2010 census, the population was 23
-
25) and 0.72 (95% CI = 0.52–0.86; n = 27), respectively. Results indicated that body size, standardized by sex, was the most important cov
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: how to use a calculator in a real situation, how to use my brain, how to use my time, how to keep moving, how to do this, how to be intentional, how to learn, how to solve a common mistake,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Llanrhydd. Llannerch contained the parishes of Llanfair D.C., Llanelidan and Derwen,<|endoftext|>Ulcers are a common scourge in horses. Gastric ulcers affect up to
-
center regularly updates its Web site with new resources about different content areas and offers poster sessions at the January and July residencies. LibGuides: Commerce Research Guide: Where can find journal ...⏎Functioning as a research and publishing network, ScienceOpen
-
for about 23 percent of the oil and 8 percent of total natural gas produced in the United States (approximately 82 percent of the oil supply comes from deepwater facilities), and the Gulf Coast region is home to almost half of
-
ography Visual Data: Photography, Film, and Video Using Documents as Data Qualitative Online Research: Using the Internet Observation and Mediated Data: An Overview PART SIX: FROM TEXT
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
- L7:911 suppresses
- L7:63 excites
- L7:65 excites
- L10:762 suppresses
- L9:1308 weakly suppresses
- L5:654 weakly excites
- L9:1033 weakly suppresses unclear: end-of-text markers and word fragments
- L7:1130 weakly excites
- L10:258 weakly excites
- L10:1015 weakly excites
- L3:1703 weakly suppresses
- L10:284 weakly suppresses
- L8:4 weakly excites
- L3:58 weakly excites
- L8:183 weakly suppresses
- L5:821 weakly excites
- L7:424 weakly suppresses
- L7:12 weakly suppresses
- L6:898 weakly excites
- L5:64 weakly excites
- L10:959 weakly suppresses
- L0:1380 weakly suppresses
- L2:1995 weakly excites
- L3:719 weakly suppresses digits inside numbers and statistics
Feeds later layers
- L13:617 excites
- L14:1620 excites
- L12:2726 excites
- L13:1242 excites
- L13:496 suppresses
- L13:2091 excites
- L13:1589 suppresses
- L14:1473 excites
- L12:190 excites
- L13:808 suppresses
- L13:1468 excites
- L14:912 suppresses
- L13:2156 excites
- L14:1533 excites
- L13:429 excites
- L13:619 excites
- L13:997 excites
- L14:398 excites
- L15:820 suppresses
- L13:2186 excites
- L12:546 excites
- L12:28 suppresses
- L15:2660 excites
- L15:150 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
- antu
- migrationBuilder
- (ierr
- kat
- rom
- 延
- uai
- 的那种
- 一面
- Hait
Suppresses
- config
- ��
- 来这里
- calc
- eldon
- icles
- ...")⏎
- andles
- anatomy
- bullet