opening parenthesis in lists and asides
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. 3 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 responds to an opening bracket '(' in the middle of a sentence, especially in lists or side notes like '(b)' or '(near-field)'. It doesn't seem to respond to the bracket itself so much as to the spot where a list item or aside begins. Turning it up gave text that did not show this pattern.
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.
-
antibodies and (b) the observation that human CD18 can form chimeric heterod
-
protected by thin double-walled carbon nanotubes. The synthesis of very long arrangements is confirmed by a combination of transmission electron microscopy, X-ray diffraction and (near-field) resonance Raman spectroscopy. Our results establish a route for the
-
buried for many Ice Age cycles, (but) when the ice advanced over this area, it was essentially frozen to the
-
electricity in schools, it is recognised that energy for other purposes such as cooking, hot water, and (in some regions) space heating, also plays an important role in improving educational quality
-
0 times faster than any other present method with comparable resolution, (such as super-resolution visible light microscopy or electron microscopy). Professor Hwu estimates that
-
digits of 2n. The number of digits can be estimated as n log102. Thus, the probability is about cxn, where c = (10/9)2 ≈ 1.2 and x = (
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 will not. I will not. I will not. I will not. I will not. I will not. I will not. I will not. I will not. I will not. I will not. I will not.
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Chegg Com⏎Derivation Of Kir<|endoftext|>The poisonous Gartraps are the bane of the desert and semi desert. They are the same colour as the sand in
-
|• Summer (DST)||UTC-5 (CDT)|⏎|GNIS feature ID||685916
-
your alcohol and caffeine consumption. Avoid excess caffeine, and no caffeine after 1pm. Aim to consume no more than 1(for females) or 2 (for males) alcoholic beverages per day. Give yourself a break from alcohol at least
-
also be empty (Brooks, 2003); complex issues are broken down into broad
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
- L0:1252 suppresses
- L2:353 weakly excites
- L2:2712 weakly excites
- L0:1798 weakly suppresses
- L1:448 weakly suppresses
- L1:1976 weakly suppresses
- L0:734 weakly suppresses
- L1:23 weakly excites
- L1:521 weakly suppresses "hand" in "on the other hand"
- L0:11 weakly excites
- L0:86 weakly excites
- L2:1784 weakly excites
- L0:1667 weakly excites
- L2:41 weakly suppresses
- L1:165 weakly excites
- L2:558 weakly excites
- L2:1126 weakly suppresses
- L0:2329 weakly excites
- L0:404 weakly excites
- L1:957 weakly suppresses
- L2:325 weakly suppresses
- L2:2828 weakly excites
- L1:499 weakly excites 'in' after success or help verbs
- L1:340 weakly excites
Feeds later layers
- L6:144 excites
- L6:468 weakly suppresses
- L6:2957 weakly suppresses
- L18:980 weakly excites opening parenthesis after a noun phrase
- L7:2907 weakly excites
- L6:243 weakly suppresses
- L4:1116 weakly excites
- L21:2576 weakly suppresses
- L24:1809 weakly suppresses
- L23:2916 weakly suppresses
- L5:2556 weakly excites
- L24:505 weakly excites
- L24:2157 weakly suppresses
- L4:797 weakly excites
- L23:2191 weakly excites
- L23:2132 weakly suppresses
- L11:50 weakly suppresses
- L5:875 weakly suppresses
- L10:2287 weakly excites
- L16:522 weakly suppresses
- L15:87 weakly suppresses
- L26:510 weakly excites
- L18:1021 weakly suppresses
- L9:2611 weakly suppresses 'is' right after 'there'
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
- deb
- бор
- DE
- .deb
- icolor
- 百家
- borrower
- ammen
- (DE
- .BL
Suppresses
- uces
- =<?=
- faces
- =min
- parameters
- ucing
- fläche
- 碣
- ean
- apologies