the letter z inside words
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 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
The neuron fires on the letter 'z' when it appears in the middle of a word, most clearly in 'unzipping'. Its weaker opposite is the letter 'm' inside words like 'hypermiling'. Turning it up or down gave repetitive or unrelated text.
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.
-
ma, gzip, compress, xz, and Izop. here we will be using the .tar utility to Unzip the files.⏎The command for unzipping the .tgz file using the .tar utility:⏎tar -xv
-
Great Migration and surpassed two million occupants by 1950.⏎http://en.wikipedia.org/wiki/Philadelphia 360-degree panorama photography by yunzeng liu.⏎360-degree panorama photography by yunzeng
-
CT and Leeds University, which examined how market forces and digital technology are changing higher education, particularly in respect of emerging models of teaching and learning.⏎Professor Laura Czerniewicz, director of UCT’s Centre for Innovation in Learning and Teaching
-
a variation of zipped files; most open-source packages use this format to zip their files. If you are new to Linux or recently encountered the problem of unzipping the .tgz file, then you don’t have to worry this guide
-
with Extract Compressed Zipped Folder will appear with the zip file path.⏎Step 3: Tap on Browse and select a destination to save the files after unzipping.⏎Step 4: Enable Show Extracted Files When Complete to see the
-
de Bèze, his enthusiasm for later kabbalistic literature clearly waned. Mercier translated almost the whole of Targum Jonathan b. Uzziel on the Prophets; and he wrote annotations to Santes *P
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:uuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuuu
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
no more forgetting distant memories. Hmm …⏎Via Sustainable Business.<|endoftext|>Word of the Year: "Hypermiling"⏎For those outside the green movement, hypermiling is defined by the word-meisters of Oxford this way: "Hyperm
-
of the waning Moon IKrishna paksha) in Vedic astrology. The phase of the Moon is calculated daily in the Vedic calendar. Krishna Janmashatmi is celebrated from moonrise to midnight on the tithi of
-
School from 1883 to 1892. She excelled in her studies: in 1885 she received the Dalrymple Victoria prize; in 1888 she won an Education Board Junior
-
hypertelorism (widely-spaced eyes); (P)ulmonary stenosis (obstruction of the normal outflow of blood from the right ventricle of the heart); (A)bnormalities of the genitals; (R
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
- L4:216 strongly excites
- L22:1766 strongly suppresses
- L16:0 strongly suppresses
- L21:196 strongly suppresses
- L21:690 strongly excites
- L19:39 strongly excites
- L16:10 suppresses
- L0:2215 suppresses
- L19:89 excites
- L2:761 suppresses
- L2:1083 excites
- L26:434 excites
- L2:1091 suppresses
- L20:2984 excites
- L2:801 suppresses
- L4:490 suppresses decade digit in years like 1930s
- L25:956 excites
- L22:271 suppresses
- L16:3 suppresses
- L16:1 suppresses
- L5:97 suppresses
- L24:980 suppresses
- L20:854 excites
- L20:961 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
- oeff
- Success
- stile
- ーム
- Authenticated
- confirmed
- جموعة
- .styleable
- ousedown
- ثل
Suppresses
- u
- uC
- uations
- an
- ução
- ynomial
- uin
- m
- uem
- uí