unclear: mixed word fragments and phrases
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. 27 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 a grab bag of word pieces and phrases with no clear shared theme. When it was pushed down, the model kept writing the word 'string'. When pushed up, it got stuck repeating words.
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.
-
be great for routine maintenance, but they may not have the knowledge to apply the correct gardening principals to your more complex gardening needs.⏎Common pruning mistake! The uniform shearing of shrubs may be more pleasing to the eye for some but, as
-
God and is profitable for doctrine, for reproof, for correction, for instruction in righteousness: That the man of God may
-
virus.⏎The hepatitis B vaccine consists of a small protein from the surface of the hepatitis B virus called the hepatitis B surface antigen
-
section 3: Extended Library Functions, Volume 2 Oracle Solaris 11.1 Information Library|⏎- load exponent of a floating point number⏎c99 [ flag... ] file... -lm [ library...
-
had been there; apparently he planned on it, but he is ill. He has spearheaded a remarkable project that has united the fields of homotopy theory, mathematical logic, and the theory of programming languages — and in the process, it
-
1 , stored in file hw3_ex1.asm , that prompts the user to type a 5-character string. The program reads the first 5 characters of that string and prints two strings. The first string contains the 5 characters entered by
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: string string string string string string string string string string string string string string string string string string string string string string string string string string string string string string string string string string string string string string string string string string string string string string string string
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Angle Sum⏎Bit to Gigabyte⏎Find the treasures in MATLAB Central and discover how the community can help you!⏎Choose a web site to get translated content where available and
-
the pagan message concealed in the thought of the contemporary Greek Platonist Georgios Gemistos Pletho and warned of the danger to orthodoxy if a Platonist like Bessarion were ever elected pope. The book precipitated
-
ak Dun v. Yukon, 2015 YKCA 18 (the “appeal decision”), the Yukon Court of Appeal unanimously allowed the Yukon government’s
-
only provides valuable insights into an organization and its customers, but it empowers teams to take action.⏎However, incomplete or inaccurate data can bring it all to a screeching halt.⏎The importance of clean data⏎Clean data — data that is
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
- L24:905 strongly suppresses
- L26:229 excites
- L23:2751 suppresses
- L23:455 suppresses
- L24:479 excites
- L22:1136 weakly excites
- L13:1469 weakly suppresses
- L22:1920 weakly suppresses
- L15:1656 weakly suppresses
- L23:1647 weakly excites
- L1:1541 weakly excites
- L21:546 weakly suppresses
- L21:225 weakly excites
- L10:886 weakly excites
- L5:216 weakly suppresses
- L9:878 weakly excites
- L1:2027 weakly excites
- L14:2520 weakly suppresses
- L17:265 weakly suppresses
- L24:3065 weakly suppresses
- L15:1203 weakly excites
- L12:1131 weakly excites
- L10:32 weakly excites
- L18:1016 weakly suppresses
Feeds later layers
None: this is the last layer.
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
- erton
- ��
- atically
- мент
- feared
- 国安
- inson
- otten
- ации
- achts
Suppresses
- (string
- string
- (vector
- string
- →string
- vector
- ,string
- vector
- (object
- ,vector