colon in a book or article title
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. 3 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 colon that separates the main title of a book, article or report from its subtitle, such as "Machine Made: Tammany Hall...". Turning it up did not produce a clear change in the model's replies.
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.
-
politicians who kept that famously well-oiled patronage machine of old working for decades at least did something to earn their graft.⏎In a new book, Machine Made: Tammany Hall and the Creation of Modern American Politics, (Norton, New
-
3th USENIX Security Symposium, 2004.⏎Garfinkel, S. and Shelat, A., "Remembrance of Data Passed: A Study of Disk Sanitization Practices," IEEE Security and Privacy, January/
-
itamins-and-nutrients/how-vitamin-c-supports-a-healthy-immune-system⏎- Kennedy, David O. “B Vitamins and the Brain: Mechanisms, Dose and Efficacy—A Review.” Nutrients,
-
Life.com and the new book Plastic-Free: How I Kicked the Plastic Habit and How You Can Too (Skyhorse).
-
on the book "Dots and Dashes: Adventures at the Dawn of the Wireless Age" is at http://vubooks
-
and editorial matter, Kerry J. Kennedy and John Chi-Kin Lee; individual chapters, the contributors.⏎|Title of host publication||Religious education in Asia: Spiritual diversity in global
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: it's not. I need to get a better answer. I need to get a better answer. I need to get a better answer. I need to get a better answer. # I need to get a better answer. Here's a
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Philadelphia, Pa: Elsevier Saunders; 2010:chap 73.⏎Karcher DS. McPherson<|endoftext|>Discrimination and disparity within correctional institutions
-
niere, at the River May (now St. John’s, Florida), in 1564. […]⏎Jacques le Moyne de Morgues (c. 1533–1588) was a
-
symptom relief, with the least restrictive diet possible.<|endoftext|>Evaluating Tagging Methods and Movement Patterns of Round Gobies⏎The round goby (Neogobius melanostomus) is an invasive fish species introduced in the St
-
Until well into the 20th century, knowledge of Jacques Le Moyne de Morgues was extremely limited,<|endoftext|>THE FATEFUL DECADE⏎THE FATEFUL
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
- L2:111 suppresses
- L2:75 weakly suppresses
- L0:769 weakly suppresses
- L0:491 weakly suppresses
- L1:988 weakly excites
- L1:1957 weakly excites
- L2:1321 weakly suppresses
- L0:1224 weakly suppresses
- L1:2128 weakly excites
- L1:1279 weakly excites
- L2:852 weakly excites
- L0:1301 weakly suppresses
- L2:483 weakly suppresses
- L2:560 weakly excites
- L2:430 weakly excites
- L0:1197 weakly suppresses
- L0:178 weakly suppresses
- L0:36 weakly suppresses
- L2:70 weakly excites
- L1:2270 weakly excites
- L2:998 weakly suppresses
- L0:869 weakly excites
- L2:1564 weakly excites
- L2:911 weakly excites
Feeds later layers
- L5:2629 weakly excites
- L13:1501 weakly suppresses
- L15:1594 weakly suppresses
- L10:2365 weakly excites
- L27:2422 weakly suppresses
- L25:1612 weakly suppresses
- L13:2036 weakly excites
- L27:1795 weakly suppresses
- L11:1721 weakly suppresses
- L26:1689 weakly suppresses
- L19:1117 weakly excites
- L4:2163 weakly suppresses
- L23:1565 weakly suppresses
- L15:1645 weakly suppresses
- L14:1921 weakly excites
- L12:440 weakly excites
- L12:228 weakly excites end-of-document token after web boilerplate
- L20:1387 weakly excites
- L23:2871 weakly suppresses
- L21:1892 weakly excites
- L16:2112 weakly suppresses
- L17:135 weakly excites
- L9:2882 weakly excites
- L8:1305 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
- rgan
- 不死
- idge
- Cas
- Cas
- otherwise
- Case
- southwest
- ensitive
- either
Suppresses
- =DB
- DBC
- "],["
- getBlock
- cular
- ',{⏎
- '],⏎
- huyết
- 搬
- ],[