'get', 'go' and 'got' after a subject
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 lights up on verbs like 'get', 'go' or 'got' that come right after a subject such as 'let's', 'it' or 'can'. When it was turned up, the model's writing drifted into repeated, rambling phrasing, so the effect on the text was not clear.
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.
-
.⏎There is another marker at this same location: MEEK/ELLIOTT THE TERRIBLE TRAIL<|endoftext|>Elephant Island Cruises⏎This island probably got its name after the number of Elephant seals sighted by Captain George Powell in
-
space and bandwidth while installing downloading and installing the packages.⏎This article will cover two methods of unzipping the .tgz files using the terminal. Let’s get into these.⏎Note: For this particular tutorial, we will use the Ubuntu
-
ator line benefits from its climate and agriculture production. Before you go further, it is good for you to enrich your agriculture production countries here. Now, let’s get started answering your question about why Southeast Asia is the largest agricultural production region.⏎Behind
-
Files?⏎We can zip and unzip files in Windows 10 easily. To zip files:⏎Step 1: Go to File Explorer and go to the files and media you want to zip.⏎Step 2: On an
-
Beach Park is identical in every way to the black-crowned night-heron overseas. Although this bird is native, perhaps it got here only recently, because it hasn't changed from its relatives elsewhere.⏎By contrast
-
babies come from? These questions populate our lives and inform our perceptions of the world. Explore these questions according to the expressions of artists from all over the world. Go beyond Western art to consider where art comes from and how it impacts all of us
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 first part of the first part of the first part of the first... - - - - - - - - - - - - - - - -22/29.22922292222
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
to protect yourself and your loved ones from its impact. So, let’s begin.⏎How Does Current Work?⏎Before going deeper into the concept of contact current, let’s take a moment to understand how current works.⏎Current is the rate at which
-
and fill it with water. Poke a small hole in the bottom of the bag at one of the bucket drain holes. Let the water drain out slowly near the trunk, but not on<|endoftext|>Electric cars may
-
Let’s see if you know any of these enormous structures.⏎Let’s check out these huge black holes in the video on page 2<|endoftext|>Ser
-
and teachers has been very positive. Program organizers report "many students say they're determined to visit Lake Mead NRA in the future."⏎It's a pretty<|endoftext|>Let’s go back in time, to Los Angeles in 1875.
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
- L2:2029 strongly excites
- L16:142 strongly excites
- L0:2471 strongly excites
- L17:360 strongly excites
- L6:944 strongly excites
- L1:24 strongly suppresses
- L17:512 strongly excites
- L6:1434 strongly suppresses
- L7:342 strongly excites
- L4:711 suppresses
- L17:399 excites
- L10:1228 suppresses
- L16:205 excites
- L16:671 excites
- L16:374 excites
- L17:517 suppresses "but" in "not only X, but" phrases
- L16:1066 suppresses
- L7:219 suppresses
- L16:353 excites
- L0:690 excites
- L2:1585 suppresses
- L16:215 suppresses
- L15:114 excites
- L17:838 suppresses
Feeds later layers
- L20:236 strongly excites
- L19:690 strongly excites
- L25:340 suppresses
- L21:2864 excites
- L27:2364 excites
- L23:2836 excites
- L19:593 suppresses
- L20:256 suppresses
- L27:2610 excites
- L19:906 suppresses
- L19:1865 suppresses
- L27:2558 excites
- L27:2235 excites
- L20:2399 suppresses
- L27:2939 excites
- L27:2566 excites capital G starting a name or word
- L24:1837 excites
- L27:2695 excites
- L26:2719 weakly excites
- L19:2330 weakly suppresses
- L19:2417 weakly suppresses
- L26:1613 weakly excites
- L19:1004 weakly suppresses
- L19:594 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
- busters
- .TabStop
- bye
- parity
- olang
- blk
- illis
- LoggerFactory
- →push
- Braz
Suppresses
- usement
- 多く
- γ
- AndGet
- uncate
- graph
- gal
- eu
- abwe