sentence-ending periods in prose
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. 14 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
The neuron lights up on full stops that close a sentence in ordinary essay-style writing, often at the end of a clause that introduces a topic. It goes quiet or negative on colons and line breaks in headings and short labels. Turning it up made the writing repetitive but did not show a clear theme.
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.
-
been thinking deeply about the alphabet—about which letter-shaped path of recovery the world will follow out of its recession. The economy could rebound sharply, tracing a V; it could trudge along the bottom a while longer before rising slowly into a U
-
, but it’s always happening.⏎Animation is no longer a novelty for web designers…it’s becoming the basis of effective interaction design.⏎Animation is change, and motion. It’s the closest we get to capturing
-
imaginative state of Greece in her best period, represented wealth, commerce, and conquest, in a greater degree, so were her arts, and with these the lyric. In her best state her nobles danced, Appius Claudius excelled,
-
up a depot for the provision of ships under the auspices of The Dutch East India Company, were of Dutch stock. Many later settlers were of German origin and after the revocation of the Edict
-
family members, or student partners.⏎- The students' stories can be compiled into a slide show or movie.<|endoftext|>Hyperion: A Fragment 1818 (1820
-
⏎to his Life he was also called Kung-tê-jih (i.e., merit-sun; in Sanskrit, Puṇyaditya?). For he was not only a philosopher, but a preacher and an organiser,
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: that the "2011,2012" "2013" "2015" "2017" "2019" "2011" "2011
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Seeking Help:⏎Ben Locke, Ph.D., a mental health expert serving at Pennsylvania State University,
-
news articles may be copied or forwarded for individual use only. No other reproduction or distribution is permitted without prior written consent.⏎Please login or Register to leave a comment.<|endoftext|>We’ve long known that racism and discrimination negatively impact the mental health and well
-
elements of a successful speech. Not only do they get to become better-informed...⏎7th - 11th Social Studies & History⏎President Obama's State of the Union Address⏎Students watch the State of the Union
-
Therefore, children may need explicit instruction in recognizing alternative interpretations in their writing. (A sample story and four figures of data are included.) (RS)⏎Publication Type: Speeches/Meeting Papers; Reports - Research⏎Education Level: N/A⏎
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
- L11:132 excites period ending an opening topic sentence
- L5:168 suppresses
- L10:612 excites full stop ending boilerplate or a title
- L6:250 excites
- L9:16 suppresses unclear: periods inside numbers and names
- L4:7 excites
- L8:541 excites
- L9:12 suppresses
- L9:2447 excites
- L9:2302 excites
- L8:740 excites
- L13:345 excites
- L12:2141 suppresses
- L10:478 suppresses
- L4:102 excites
- L3:259 suppresses
- L11:1194 excites
- L9:819 suppresses
- L5:155 weakly suppresses
- L10:945 weakly excites
- L4:51 weakly excites
- L11:1272 weakly excites unclear: sentence ends in factual prose
- L4:562 weakly suppresses blank slot between a date's day and year
- L2:9 weakly suppresses sentence-final periods vs dots in code
Feeds later layers
- L17:990 excites
- L16:1545 excites
- L17:1155 excites opening parenthesis after a sentence
- L19:94 suppresses
- L15:403 excites period inside an abbreviation
- L19:2097 excites sentence-end punctuation in speech
- L16:489 suppresses
- L17:1105 excites
- L16:894 excites full stops ending informational sentences
- L15:1207 excites
- L21:524 excites
- L17:1772 excites
- L17:231 suppresses
- L16:949 excites
- L15:21 suppresses
- L17:347 excites full stops ending informative sentences
- L17:1044 excites
- L19:2697 excites
- L19:909 excites
- L18:1465 excites semicolon joining clauses
- L16:436 excites
- L19:433 excites
- L16:1608 suppresses
- L16:797 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
- 正当
- 気が
- I
- This
- 在这个
- A
- trl
- 据了解
- itu
- 随着
Suppresses
- %",⏎
- ;",⏎
- !",⏎
- :",⏎
- orns
- $",
- /she
- /",⏎
- \",\
- '{$