the spot right before a number
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. 27 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 position just before a number, such as an age, a year, a dosage, or a count, so the model is about to write digits. Turning it up made the model produce strings of digits, which matches what it seems to be preparing for. Its negative side fires on the digits inside a number.
Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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.
-
a young age).⏎Guidelines differ for men at average risk. Some recommend an initial PSA (Prostate Specific Antigen) blood test and clinical examination at age 40, and others recommend starting at age 50. In general,
-
with an abiding desire for racial equity.⏎Higher education has never been a colorblind meritocracy. Until the implementation of equal opportunity and affirmative action programs in the 1960s, most colleges and universities were historically white institutions. Schools
-
20 were women finally guaranteed the right to vote.”⏎On August 24, 1972, Nixon said:⏎“For the first time in the 195 year history of this country, men and women 18 to
-
0, in addition to regular self-exams and exams by your doctor or other health care provider.⏎- Colon and rectal cancer: If you are over age 50, talk to your doctor about getting one or more of several screening tests
-
see DESCRIPTION]. If either of these conditions exists, the vaccine should not be administered.⏎Between uses, return the multidose vial to the recommended storage conditions between 2° and
-
-Gammon reminds us that while it could be harmful, we should remember that the storm itself could have more dire consequences.⏎"If you're looking at a Category 4 or
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:
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
samples in one full cycle, you will see that there are 45. Divide 44,100 by 45 and you get 980. (I use 980 hertz here rather than
-
water factors:⏎Fat: 9 cal/g Carb: 4 cal/g Protein: 4 cal/g⏎- Percent Daily Values are based on a 2,
-
now accepted defeat.⏎The brief, fragmentary Book <|endoftext|>Welfare state in capitalism⏎Since the eruption of the sub-prime mortgage crises of 2007 to 2009, there have been hot debates on how the
-
in pyrotechny, imparting an intense red color to flames.⏎- n. A metallic chemical element (symbol Sr) with an atomic number of 38.⏎GNU Webster's 1913⏎- n. (Chem
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
- L21:339 strongly excites spot just before a year
- L25:146 strongly excites
- L26:216 strongly suppresses blank slot before a year or number
- L25:854 suppresses
- L25:182 excites
- L22:2275 excites
- L19:581 suppresses blank slot before a number
- L24:197 excites
- L23:77 excites the space right before a number
- L26:121 suppresses
- L25:3007 suppresses the zero digit in decimals
- L22:743 suppresses
- L25:33 excites spot where a number is due
- L18:377 excites number right after 'about' or 'less than'
- L22:372 excites empty spot before a number or "3D"
- L20:962 excites decimal points and dots in numbers
- L21:395 excites
- L23:199 suppresses
- L26:868 suppresses
- L4:53 excites blank slot before a number
- L26:15 suppresses
- L26:103 suppresses blank space before a number
- L26:1551 excites
- L25:232 suppresses
Feeds later layers
None: this is the last layer.
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
- 5
- 3
- 1
- 2
- 4
- 0
- 8
- 9
- 6
- 7
Suppresses
- ⏎
- ⏎
- ⏎
- bones
- ”?
- ⏎⏎
- ”:
- ⏎
- '?
- ⏎