full stop ending boilerplate or a title
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. 10 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 reacts to the full stop that closes a short line, such as the dot in 'This preview shows pages 1 to 2.' or the end of a title. When it was turned up, the model's text did not show a clear theme.
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.
-
is not a viable option, as it is a highly-energy intensive process, and half or more of the water produced<|endoftext|>This preview shows pages 1–2. Sign up to view the full content.⏎
-
caused<|endoftext|>This preview shows pages 1–2. Sign up to view the full content.⏎This preview has intentionally blurred sections. Sign
-
conservative views that are particularly detrimental for women.⏎But nothing — not even a conservative culture or fear of the Taliban<|endoftext|>This preview shows pages 1–2. Sign up to view the full content.⏎This preview has intentionally blurred sections. Sign
-
SciPy or R programs.<|endoftext|>by Roman Davidov⏎(revisited from the June 2003 issue of Meteorite Times)⏎Zhamanshin. This word translated from Kazakh language like “bad soil” or “land without
-
Research Laboratory, Pittsburgh, PA 15236⏎Other Occupational Concerns⏎Longwall USA, International Exhibition & Conference<|endoftext|>Hypertext Markup Language. The standard language used to create Web pages. Originally based on SGML, note
-
in synthesis Home⏎PDF. Rich HTML. We have demonstrated purely mechanochemical hydrogenation of MgB2
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: about in 1995-2000 in the first year of work at the 2000 research in the 1900s in the 1990s in the 190
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
scent of baby shampoo under your nose? Most of us think such moments are gone after our snuggler learns to read. But they don’t have to be! Save the ritual, and you can help her become an even better reader than she
-
ery, and power politics—all wrapped up in a story befitting a good dime store novel. And the proverbial plot thickened when computer scientists got involved—at that point, things got downright ugly. As painful as it was for the computer
-
over?⏎Quite simply, if you actually look at the broad measures of aggregate economic activity that define recessions and recoveries, they don’t look like those letters—at best they would be Procrustean fits. For example, the Great Depression
-
because dopamine connects neurons, which helps you recognize everything associated with a rewarding experience. The tribe is thrilled by the fish feast—for a few days. But then the dopamine sags. So you go out for
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
- L6:250 excites
- L4:7 excites
- L9:819 suppresses
- L4:3 excites commas after citation fields
- L9:458 excites
- L5:168 suppresses
- L4:1546 suppresses
- L5:974 weakly excites
- L9:1802 weakly excites
- L8:517 weakly suppresses
- L5:1653 weakly excites
- L3:487 weakly suppresses
- L9:1113 weakly suppresses
- L9:16 weakly suppresses unclear: periods inside numbers and names
- L7:948 weakly suppresses
- L5:454 weakly suppresses
- L6:2621 weakly suppresses
- L3:640 weakly suppresses
- L8:93 weakly suppresses period after a middle initial
- L1:2157 weakly excites
- L8:2184 weakly excites
- L8:573 weakly excites
- L8:1578 weakly suppresses
- L9:121 weakly suppresses
Feeds later layers
- L16:1112 excites
- L14:155 excites sentence-ending periods in prose
- L17:1772 excites
- L17:1105 excites
- L21:1086 excites end of sentence before a claim or statistic
- L14:153 excites
- L13:9 suppresses
- L17:990 excites
- L16:241 suppresses
- L11:1272 excites unclear: sentence ends in factual prose
- L24:148 excites
- L14:409 suppresses
- L16:436 excites
- L13:345 excites
- L17:1770 excites
- L11:132 excites period ending an opening topic sentence
- L19:2097 excites sentence-end punctuation in speech
- L12:771 excites
- L16:2535 excites
- L17:839 excites
- L17:2613 excites
- L19:2614 excites
- L16:1545 excites
- L16:1082 excites comma linking two clauses after 'and'
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
- elay
- roker
- whereas
- enez
- кроме
- bra
- 而在
- ollo
- submit
- 有条件
Suppresses
- t
- political
- ander
- orestation
- 岚
- oat
- Meal
- /L
- is
- -te