spot where a number comes next
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. 15 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 at the point just before a number is expected, such as the gap in a zip code, a confidence interval, or a number sequence. It is strongly negative on the middle digits of years like 1856 or 1916. When turned up, the model's text did not show a 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.
-
—and it can pass in a flash if you don't have a method for attacking it.<|endoftext|>304 North Cardinal St.⏎Dorchester Center, MA 02124⏎When you’re manually operating a boat, the act
-
015 The University of North Carolina Press⏎116 South Boundary Street, Chapel Hill, NC 27514-3808⏎How to Order | Make
-
1, 1, 2, 3, 5, 8, 13, 21, 34, 55… each number is the sum ( + ) of its two previous numbers
-
996) and abdominal pain (p = 0.018, OR = 0.815, 95% CI = 0.68
-
⏎For more info about recovery and treatment strategy and options, click here⏎626 S. Ferguson Ave. Suite 5⏎Bozeman, MT 59
-
tau levels at 24-48 hours post injury (mean difference –0.51 pg/mL, 95% confidence interval, –0.88 to –0.14; P = .008)
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: I have a function that is a different version of the same function in the same family of functions in a different version of the solution, I have a function in the same family of a different version of the same function of the same family in
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
, of the University of Birmingham, and colleagues concluded.⏎Disclosure: The researchers report no relevant financial disclosures.<|endoftext|>You are here⏎Edward Elgar (1857-1934) was 50 years old before he began
-
process that people go through when formulating solutions to problems.⏎Herbert Simon was born in Milwaukee, Wisconsin, on June 15th, 1916. He went to the Milwaukee public school system,
-
Curtis, George Ticknor, American writer on legal topics, was born at Watertown, Mass., Nov. 28, 1812; and died at New York,
-
. A statue of him was erected to commemorate him in London's Parliament Square.⏎Early life and education⏎He was born on 24 May 1870, at the family farm, Bovenplaats, near Malmes
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
- L14:542 excites
- L11:2554 weakly excites
- L14:2846 weakly excites
- L4:2337 weakly suppresses
- L14:2677 weakly excites
- L1:436 weakly suppresses
- L7:1333 weakly excites
- L6:2866 weakly excites
- L2:134 weakly excites
- L8:421 weakly excites
- L9:1881 weakly excites
- L14:1627 weakly excites
- L8:3020 weakly excites
- L5:2943 weakly excites
- L12:107 weakly excites
- L14:1700 weakly excites
- L6:2093 weakly excites
- L14:1829 weakly suppresses
- L13:2285 weakly suppresses
- L10:288 weakly excites
- L7:1609 weakly excites
- L6:1320 weakly excites
- L2:2639 weakly suppresses
- L12:697 weakly suppresses
Feeds later layers
- L26:316 weakly excites
- L20:1924 weakly suppresses
- L17:351 weakly excites
- L18:544 weakly excites
- L17:2980 weakly excites
- L19:672 weakly excites
- L16:17 weakly excites
- L23:1829 weakly suppresses
- L21:349 weakly excites
- L17:2496 weakly suppresses
- L19:1523 weakly suppresses
- L17:119 weakly excites
- L19:2376 weakly suppresses
- L16:135 weakly excites
- L17:328 weakly excites
- L20:929 weakly excites
- L23:3055 weakly suppresses
- L19:673 weakly excites
- L21:545 weakly excites
- L24:8 weakly suppresses
- L23:194 weakly excites
- L16:992 weakly excites
- L17:784 weakly excites
- L22:856 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
- 邮箱
- 峁
- emperature
- 둘
- çeş
- turned
- าง
- 对方
- paged
- 율
Suppresses
- عني
- cı
- Corps
- uras
- 蒟
- hypoc
- mek
- UIS
- metaph
- uição