dotted numbers and key parts
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. 9 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 reacts to the dots and digits inside version numbers and IP-style addresses, like 22.214.171.124, and to the word "private" in talk about encryption keys. It goes strongly negative on words about password characters, such as "special" and "letters". Turning it up gave replies that were unrelated, so the link to its trigger is unclear.
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.
-
68k-palmos-coff-gcc -v should report its version number to be (at time of writing) 22.214.171.124-kgpd-07109
-
the 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
-
ated by Maria Ciaccio⏎©2018 Chicago Botanic Garden and my.chicagobotanic.org<|endoftext|>Hello. Sign In⏎Please enter a valid city, state or zip code⏎With the Food and Drug Administration (FDA
-
as being a List of Users.⏎Sometimes, it’s also handy to define a list of multiple Entities and Structures.⏎Until 184.108.40.2
-
uler or calipers with mm divisions⏎- Paper towel⏎- Microscope slide and cover slip⏎AQA: Topic 184.108.40.206 and 220.1<|endoftext|>Re export
-
, 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
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 list of characters that contains a mix of numbers, symbols, and letters. This is a password... That's a password... that includes numbers, symbols, and letters. That's a password... that includes numbers, symbols,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
file represents a sequence of characters (letters, numbers or special characters) that are stored on a personal computer, accessible only to the PC owner
-
Even the simplest version is a fairly difficult password to crack. If you’re really security conscious you can make passwords stronger by:⏎- Use both upper and lower case letters⏎- Use more special characters⏎- Make the way you combine them more complicated
-
for Amazon, John Bob Smith could use JBS10AMA. To make a password even more secure, or applicable for services that require other characters, add special characters around it, like # JBS10AMA#.⏎Here are some
-
., Either-OR⏎The value has to be one of the following-⏎- Statment 1⏎- Statement 2⏎Regex for checking if its a valid Either of Two words i.e., Either-OR⏎Test string examples for the
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
- L6:1361 suppresses
- L6:1307 weakly excites
- L5:2 weakly suppresses
- L8:5 weakly suppresses
- L8:1757 weakly suppresses
- L7:122 weakly excites
- L4:2725 weakly excites
- L0:279 weakly suppresses
- L7:174 weakly suppresses
- L6:7 weakly suppresses
- L7:2940 weakly excites
- L7:2605 weakly excites
- L0:2363 weakly suppresses
- L1:2725 weakly suppresses
- L4:2463 weakly excites
- L0:2528 weakly suppresses
- L6:2767 weakly excites
- L2:2827 weakly excites
- L8:632 weakly suppresses
- L2:2684 weakly suppresses
- L4:2004 weakly excites
- L8:12 weakly excites
- L7:2664 weakly suppresses
- L1:333 weakly suppresses
Feeds later layers
- L20:2760 suppresses
- L21:1220 suppresses
- L16:2197 suppresses
- L11:129 excites
- L15:352 weakly suppresses
- L10:628 weakly suppresses
- L21:2313 weakly excites
- L20:1577 weakly suppresses
- L11:1056 weakly suppresses
- L17:2246 weakly suppresses
- L14:1664 weakly suppresses
- L12:2967 weakly suppresses
- L22:1577 weakly excites
- L17:319 weakly excites
- L15:1910 weakly suppresses
- L10:2353 weakly suppresses
- L26:2934 weakly suppresses
- L27:2705 weakly suppresses
- L10:2551 weakly excites
- L17:2884 weakly suppresses
- L12:2303 weakly suppresses
- L20:2927 weakly suppresses
- L15:781 weakly suppresses
- L21:1383 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
- tribution
- 钊
- войны
- favor
- favor
- quarters
- _ctor
- cki
- 安置
- indifference
Suppresses
- validators
- validators
- validator
- 的男人
- validator
- ipa
- 合格
- 色情
- 手机号
- validate