comma after 'though' mid-sentence
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. 5 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 fires on the comma that comes right after the word 'though' when it sits in the middle of a sentence, as in 'the main city, though, ...'. Turning it up did not produce comma-heavy text; the output wandered into unrelated topics.
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.
-
start of the 1st century A.D., when the main city of Populonia was mysteriously abandoned. People continued to live at the villa, though, until the end of the 5th century, so it is one of the
-
the curves, though, when your spine is seen from the rear, it looks like it’s running straight
-
called ruby-throats, since the males share the same red gorget as their East Coast cousins. The predominant species of hummingbirds in our valley, though, is the broad-tail. These birds look similar to the ruby-throats
-
underestimated, an error that has in effect enabled others to have a stronger claim on their wealth creation role.⏎It is hard to make the pitch for government, though,
-
whether the health benefits of chocolate outweigh the additional calories consumed. Don’t be concerned about eating the occasional chocolate bar. If you are trying to lose weight, though, eating more fruits and vegetables is a better option.<|endoftext|>MLA style for documentation
-
habitat, given the unpredictable nature of the cone crop. Forests that also have at least some areas of pines, Norway Sprue or Larch, though, can sustain large populations. Some Red Squirrels do fall prey to Pine Mart
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 been using this website to find the most recent of the week: I have been told by the teacher of this week: I have been given this to you: I have been given this to you: I have been given this to
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
assets) depending on which way the decision goes.⏎Those are the ‘easy’ situations<|endoftext|>You hear it all the time, “don’t write your picture book in rhyme.” That’s because
-
awesome) discovery. Without knowing too much about it, I’m sure that there will be a raft of problems involving
-
alin. Amygdalin is a chemical ingredient found in the pits of many fruits, raw nuts, and plants (see Question 1).⏎- Hydrogen cyanide is thought to be the main anticancer compound
-
“, and others in my awesome Edumatch Voxer group, told me about BreakOut EDU which is basically the same thing as an Escape Room. Best of all, they have resources, templates, and ready-made games which you or your
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
- L0:2311 weakly suppresses
- L4:813 weakly suppresses
- L1:2502 weakly excites
- L1:1693 weakly suppresses
- L1:969 weakly excites
- L2:1996 weakly suppresses
- L3:2664 weakly suppresses
- L2:219 weakly excites
- L1:1712 weakly excites
- L1:1331 weakly suppresses
- L1:2778 weakly excites
- L1:1559 weakly excites
- L0:319 weakly suppresses blank space after a label colon
- L4:2989 weakly suppresses
- L2:1175 weakly excites
- L3:369 weakly excites
- L3:1966 weakly excites
- L1:188 weakly suppresses
- L4:2357 weakly excites
- L4:594 weakly suppresses
- L1:2646 weakly suppresses
- L3:203 weakly excites
- L3:1262 weakly excites
- L1:1992 weakly suppresses
Feeds later layers
- L14:955 weakly suppresses
- L16:2988 weakly excites
- L25:856 weakly excites
- L19:1431 weakly suppresses
- L13:1355 weakly excites
- L6:70 weakly excites
- L23:1666 weakly suppresses
- L20:1826 weakly suppresses
- L8:781 weakly suppresses
- L8:299 weakly suppresses
- L12:2276 weakly excites
- L14:134 weakly suppresses spot right before a citation year
- L12:2425 weakly suppresses
- L14:647 weakly excites
- L10:838 weakly suppresses comma in a list of paired nouns
- L25:66 weakly excites
- L25:2406 weakly suppresses
- L19:369 weakly excites
- L7:2608 weakly suppresses
- L13:161 weakly suppresses
- L10:2712 weakly excites
- L26:1969 weakly excites
- L11:56 weakly excites
- L6:1073 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
- uster
- meet
- toi
- gen
- 胆
- ánh
- widths
- get
- issues
- args
Suppresses
- candidacy
- 的时间里
- MARY
- 儿
- 了自己的
- eman
- 兒
- ikal
- 写下
- 了过来