position 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. 18 of Qwen3-0.6B. Its strongest snippets are fairly consistent: several of them 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 tends to fire on the spot just before a number, such as a blank where a figure comes next, or on a full stop before a name's initials. Turning it up produced rambling text that did not clearly relate to numbers.
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.
-
of approximately 1.7 million kilometers (1.1 million miles) from Prometheus and at a Sun-Prometheus-spacecraft, or phase, angle of 34 degrees. Image scale is 10 kilometers (6 miles) per
-
= vsin(theta) Immediately after take off : vx = 2cos25 = 1.8126 vy = 2sin25 = 0
-
ant.⏎Remove dried stems only after foliage dies back.⏎Rainy Side Notes⏎Lilium 'White Stargazer' is the most popular of the white upright lilies and
-
|⏎This species's range and population are probably declining as a result of habitat loss. It therefore qualifies as Endangered.⏎|Range Description:||Pyrrhura viridicata occurs only in the Sierra Nevada de Santa Marta
-
One Original Church⏎by the Most. Rev. Sarhad Jammo, Ph.⏎Ethnicity, Culture and Religion⏎Christian
-
oltage AC power supply⏎- Two capacitors, 0.1 µF each, non-polarized (Radio Shack catalog # 272-135)⏎- Two 27 kΩ resist
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: the [401 Pill" will be a way to be... 401 Pill is a 401 penny contribution, and the company's 2025-401PP is 521A-
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
05-2009 / Preservative-Free Flu Shots⏎03-04-2009 / Flu Strains Developing Resistance to Tamiflu⏎03-03-2009 / The History Of
-
, which specifies where the runtime can locate an assembly (for more information, see Specifying an Assembly's Location).⏎<codeBase version="188.8.131.52" href="http://www.litware
-
5⏎Paperback ISBN: 9781503603844⏎Ebook ISBN: 9781503604513⏎Request Review/Desk/Examination Copy
-
He was recently given the honor of asking Edward Snowden the first question at his SXSW appearance where he talked about Internet privacy.<|endoftext|>Disclosure :: This post is sponsored by IDEA Public Schools. When we published our first post on IDEA Public Schools we received
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
- L17:2257 suppresses
- L17:1336 suppresses
- L14:1796 suppresses
- L14:2157 suppresses
- L15:2509 suppresses
- L15:1873 weakly excites
- L12:398 weakly suppresses
- L15:777 weakly excites
- L15:949 weakly suppresses
- L9:1402 weakly suppresses
- L17:2635 weakly excites
- L16:1387 weakly suppresses
- L14:2037 weakly suppresses
- L10:2484 weakly suppresses
- L17:2621 weakly excites
- L2:103 weakly excites
- L10:2317 weakly excites
- L8:17 weakly suppresses
- L14:1436 weakly suppresses 'and' linking a role to a name
- L16:2217 weakly suppresses
- L17:2287 weakly excites
- L17:1860 weakly suppresses
- L13:1483 weakly excites
- L13:2256 weakly suppresses
Feeds later layers
- L19:2411 suppresses
- L19:929 suppresses
- L20:1174 suppresses
- L19:1607 suppresses
- L25:2383 weakly suppresses
- L19:873 weakly suppresses
- L24:710 weakly suppresses
- L19:2780 weakly excites
- L21:2847 weakly excites
- L20:38 weakly excites
- L20:191 weakly suppresses
- L19:1844 weakly suppresses
- L20:1361 weakly suppresses
- L25:1421 weakly suppresses
- L20:1232 weakly excites
- L27:2073 weakly suppresses
- L23:487 weakly suppresses
- L23:360 weakly suppresses
- L27:2694 weakly suppresses
- L20:1572 weakly suppresses
- L21:2948 weakly suppresses
- L19:586 weakly suppresses
- L25:531 weakly suppresses
- L21:557 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
- 徒
- 涿
- addObject
- trusting
- 貌似
- ello
- smouth
- 嘴
- stressing
- persist
Suppresses
- .spi
- ",[
- ponse
- theano
- łe
- ethoven
- áo
- atonin
- حال
- bote