digits inside citation numbers
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. 21 of Qwen3-0.6B. Its strongest snippets are mixed: they peak on different tokens. 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 single digits inside reference details like volume numbers, page ranges, dates and DOI codes, such as the 4 in 2010/04/19 or the 5 in 524. Steering was mixed, with number sequences showing up when it was turned up.
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.
-
+ 1 = 2. We then get the answer 524 Can we simplify? Yes, that's possible. 524 = 512.⏎The answer is 214 + 314 =
-
foods, so follow these tips to shake your salt habit:⏎*http://www.washingtonpost.com/wp-dyn/content/article/2010/04/19/AR
-
and articles⏎James Behuniak Jr (2011). Naturalizing Mencius. Philosophy East and West 61 (3):492-515.⏎David E. Soles (1999).
-
, CT scan, rule, minor head injury⏎Ann Afr Surg. 2018;15(2):57-61 DOI:http://dx.doi.org/10.43
-
, and it can exceed MRI when assessing for partial tendon tears because of its ability to perform dynamic testing. (Eur Radiol 2012;22:1140.) Ultrasonography also has been shown to be
-
native and non‐native plants to disturbance and herbivory in mountain environments. Journal of Biogeography, 48(7), 1594-1605 https://doi.org/10.111
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 to do some changes to the way the database is structured to make it easier for the users to use. To do so, I will make some changes to the database structure, and I will make some changes to the database to be
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
with a while loop:⏎given_arr = [1, 2, 3, 4, 5] i = len(given_arr) - 1 while(i >= 0): print(given_arr[i]) i = i -
-
that have a high pH value (10 to 14), which can be quite caustic. The use of acidic products
-
it that way. That just means that I have to make sure that I use gravity as a negative number, since gravity always acts down.⏎vf = vi + at⏎= 56.3m/s + (-9.81m
-
perpendicular.⏎Let's see how to identify parallel and perpendicular lines.⏎Let's take some examples to understand this.⏎Considering two equations:⏎y = 2x +3 and y = 2x+5⏎On comparing y = 2x
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
- L19:232 suppresses
- L19:1549 excites
- L18:100 weakly suppresses
- L18:2078 weakly excites
- L20:737 weakly suppresses
- L16:743 weakly excites
- L19:2083 weakly excites
- L20:2813 weakly excites
- L8:525 weakly excites
- L11:249 weakly excites
- L12:96 weakly suppresses
- L17:1152 weakly excites
- L16:325 weakly suppresses
- L18:1132 weakly suppresses
- L9:31 weakly excites
- L16:2660 weakly excites
- L12:2336 weakly suppresses
- L20:80 weakly excites
- L17:2110 weakly suppresses
- L18:2056 weakly excites
- L0:1110 weakly suppresses
- L3:1550 weakly suppresses
- L0:1980 weakly excites
- L20:156 weakly suppresses
Feeds later layers
- L22:50 weakly excites
- L23:559 weakly excites
- L26:403 weakly excites word starts like cr, fr, kr
- L27:744 weakly excites
- L26:235 weakly excites
- L23:1358 weakly excites
- L25:1030 weakly excites
- L23:1237 weakly excites
- L24:1695 weakly suppresses
- L24:1542 weakly excites
- L23:507 weakly excites
- L24:173 weakly excites
- L27:2198 weakly excites
- L24:673 weakly excites
- L26:32 weakly suppresses
- L25:1577 weakly excites
- L27:789 weakly excites
- L24:1394 weakly excites
- L22:2439 weakly excites
- L24:1878 weakly excites
- L26:316 weakly excites
- L23:1232 weakly suppresses
- L24:539 weakly excites
- L25:2689 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
- (
- send
- .Http
- (Http
- osen
- 'https
- "https
- [[
- Http
- ió
Suppresses
- ניים
- тра
- emode
- rians
- ration
- ăn
- Directive
- oji
- 简化
- ☟