dots inside version and section 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. 11 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 lights up on the dot inside numbers like version '0.5.0', section '2.2', or IP-style '188.8.131.52'. Turning it up made the model produce strings of '1.1.1.1' patterns, which fits the same idea.
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.
-
1 by Michael Faraday, and share several key functional parts.⏎- 1 Power transformers⏎- 1.1 Laminated core⏎- 1.2 Toroidal⏎- 1.3 Autotransformer⏎-
-
for-mation of the dissimilar metal joint and the measurement of joint tensile strength⏎Welding Method 6, Paragraph 188.8.131.52 (Paraphrased, still in public comment
-
grow a new Monstera plant effectively:-⏎- 2.1 Step 1: Finding the right place to grow a new plant:⏎- 2.2 Step 2: Search for the right node or petiole:⏎-
-
<|endoftext|>Mercury is used in thermometers, fluorescent light bulbs, and some electrical switches.⏎"Mercury" is in the HS Code "2805.40"⏎- - Mercury⏎
-
Figure 3-1).⏎Extension modules play an active role in all these aspects of the Apache server's⏎Figure 3-1. The HTTP transaction consists⏎of a URI request from the browser to
-
the hood. If the face velocity becomes excessive, the facility engineers should make adjustments or corrections.⏎For hoods without face velocity controls (see section 8.C.6.
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 don't know," you say, "What is the meaning of 'not work for you'?" So I'm wondering, what is the meaning of 'I'?" in the context of the word 'today'?" So I
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
specific section of an organism’s genome, causing the organism to behave differently and to pass this difference to all their offspring. CRISPR-Cas9 gene drives do this cut-copy-paste during the reproductive cycle to make sure
-
to call these ‘selfish’ elements ‘gene drives’ is not accurate. Though they are to an extent based on naturally occurring systems, CRISPR-Cas9 and/or other engineered gene drives are very different than these naturally occurring selfish genes
-
1: Download Drupal⏎First download the Drupal using following command:⏎# wget http://ftp.drupal.org/files/projects/drupal-7.33.zip⏎Step 2: Extract Compressed Drupal⏎
-
means that the files are stored in a compressed format.⏎For Python Implementation⏎You can use OpenCV to manipulate the images.⏎pip3 install opencv-python numpy⏎Open up a new Python file and follow along:⏎import numpy as np
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
- L8:1107 excites
- L10:2461 excites
- L5:475 weakly excites
- L3:1831 weakly suppresses
- L5:2928 weakly excites
- L3:1427 weakly excites
- L10:3060 weakly suppresses
- L6:1087 weakly suppresses
- L3:723 weakly excites
- L9:1587 weakly excites
- L5:885 weakly excites
- L5:3016 weakly suppresses
- L8:1618 weakly suppresses
- L9:2503 weakly excites
- L7:776 weakly suppresses
- L10:623 weakly excites
- L8:1672 weakly excites
- L10:1516 weakly excites
- L10:1510 weakly excites
- L9:1658 weakly excites
- L6:579 weakly excites
- L9:95 weakly suppresses
- L0:66 weakly suppresses
- L7:552 weakly suppresses
Feeds later layers
- L16:372 excites
- L19:298 excites
- L14:1265 excites
- L13:1083 excites
- L21:395 excites
- L13:2041 excites
- L12:2379 excites period or digit inside section numbers
- L17:1806 excites
- L13:1426 excites
- L17:870 excites
- L20:2450 weakly suppresses
- L14:2346 weakly excites
- L13:2749 weakly excites
- L13:727 weakly excites
- L14:2044 weakly suppresses
- L16:1800 weakly excites
- L15:1255 weakly excites
- L22:2931 weakly excites
- L24:2389 weakly excites
- L23:1436 weakly excites
- L19:728 weakly excites
- L26:316 weakly excites
- L13:1772 weakly excites
- L21:175 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
- icles
- Ivanka
- tep
- nonetheless
- appendix
- icle
- acknow
- subsection
- enda
- ément
Suppresses
- ere
- طلاق
- ps
- pee
- ím
- 负
- ze
- igest
- public
- ill