blank slots to fill in
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. 7 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 blank lines and gaps that someone is meant to fill in, like '___' in a worksheet or '_____' in a sentence. Some of its top hits are not blanks at all, so the pattern is only partly clear. Turning it up or down gave writing that did not reflect the blanks.
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.
-
-Davis. Demography, July 2, 2014. DOI: 10.1007/s13524-014-0308-7<|endoftext|>Designs and simulation results
-
. Its structure is similar to that of the ___2___ and national governments, with an elected mayor as chief of the ___3___ and an elected council that ___4___ the various neighborhoods forming the ___5___ . The ___6___ appoint
-
My students ran into several technical challenges; (1) creating a robust shell with adequate wall thickness, (2) how to integrated the speaker grille into the plaster form, and (3) creating a suitable mold for the plaster that could could be
-
budget. The council passes city ordinances, sets the tax rate on property, and apportions money among the various___9___. As cities have grown, ____10___ have usually come to represent more than a single neighborhood.⏎3.
-
in too little bonding performance of the slurry, and the slurry will not be able to adhere tightly to the foamed plastic net. After demolding, it will have no strength and will easily disassemble. Hang up even if
-
No! Your flexible brain is so good at making sense of patterns that it can see two images in one image.⏎The image at left is both a __________________ and __________________.⏎Your brain also solves mysteries in language. Imagine you find the cryptic note
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: "If there is a pattern, and you compare me to, "and if that pattern is true, then there you go, I think the top line of the article, you say, "and if you would say, "take a
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
*) memchr (str, 'p', strlen(str));⏎printf ("'p' found at position %d.\n", pch-str+1);⏎printf ("'p' not found
-
will present new data concerning a stereo photography studio that Dalmau opened in 1854.⏎Download data is not yet available.⏎How to Cite⏎Cuenca, C. (2019). The The “Animated
-
fanlike diagram, groups represented by rays that _______________ to the edge of the fan evolved more recently.⏎When organims are classified
-
successfully interbreed belong to the same ___________.⏎In the scientific name of the white oak, Quercus alba, Quercus is the __________________.⏎Two groups of organisms that are farther from each other on a cladogram share
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
- L6:1039 suppresses
- L6:766 excites
- L3:1613 weakly suppresses
- L6:478 weakly excites
- L3:533 weakly excites
- L6:540 weakly excites
- L6:8 weakly suppresses
- L3:2656 weakly excites
- L2:1363 weakly suppresses
- L4:788 weakly suppresses
- L0:1915 weakly excites
- L6:2874 weakly excites
- L4:2359 weakly excites
- L6:13 weakly excites
- L3:408 weakly suppresses
- L3:42 weakly excites
- L5:91 weakly excites
- L6:780 weakly excites
- L3:372 weakly suppresses
- L5:2529 weakly suppresses
- L1:1494 weakly excites
- L4:316 weakly suppresses
- L0:1464 weakly suppresses
- L6:1235 weakly excites
Feeds later layers
- L9:1449 excites
- L8:1228 weakly suppresses
- L14:2840 weakly excites
- L24:462 weakly suppresses
- L24:55 weakly suppresses
- L8:1682 weakly suppresses
- L14:766 weakly suppresses
- L11:1440 weakly excites
- L18:1683 weakly suppresses
- L17:2514 weakly suppresses
- L8:100 weakly excites
- L27:1801 weakly excites
- L13:693 weakly excites
- L13:2855 weakly excites
- L8:2598 weakly excites
- L22:1947 weakly suppresses
- L18:802 weakly excites line break after list items or paragraphs
- L22:1384 weakly suppresses
- L14:1983 weakly excites
- L9:2687 weakly suppresses
- L23:1665 weakly suppresses
- L17:2278 weakly suppresses
- L21:2023 weakly excites
- L8:1676 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
- arehouse
- -choice
- heit
- 平台
- 平台上
- Onc
- ⏎⏎
- 决
- #⏎⏎
- "]),⏎
Suppresses
- iscrim
- ände
- rist
- setInput
- änd
- istically
- setAddress
- IRST
- ルド
- heraus