commas that lead into a continuation
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. 20 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 a comma that comes after a finished clause, where the sentence is about to add more, like "...difficult to distinguish" or "...including the stomach." It dips on commas inside a list or title. Turning it up made the model repeat words like "which" and "as."
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.
-
at least the most common variety of obsidian). However, sometimes obsidian isn’t as readily identifiable as you might think. It can grade into other rock types, making it difficult to distinguish.⏎Perhaps you have found a less common variety that doesn
-
innovative drugs have been introduced, from Pfizer's (<|endoftext|>Information identified as archived on the Web is for reference,
-
have yielded highly successful results, as revealed by Yang. These triumphs pave the way for its maiden voyages
-
refer patients to a surgeon when surgical treatment is needed. Many surgeons diagnose and treat disease through nonsurgical methods, but it is usually in an area of medical specialization in which they are considered experts
-
to its Cold War importance, but it is yet underappreciated by contemporary policymakers. NATO faces threats from all directions, so how can the Alliance also pay attention to the North Atlantic? How can it
-
technical skills are required only in the IT sector, or in office-setting jobs. But that’s not true. Some level of technical
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: , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , ,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
of galaxy clusters; they are among the biggest coherent structures in the Universe (what’s bigger? Superclusters, which are clusters of clusters!). Imagine a collection of, oh, say, a thousand galaxies the size of the Milky Way. The sky
-
free()⏎- Automatic variables left on the stack of languages like C or garbage collected in languages like Java.⏎Byers, Simon. Scalable Exploitation of, and Responses to Information Leakage Through Hidden Data in Published Documents, AT&T Research,
-
posited answers to this question, whether directly - in the case of Plato and Plotinius - or through story - in the case of Homer.⏎Plotinus, having been a proponent of Platonism, naturally concurs with Plato,
-
. Think of a world where you could selectively download your life’s memories- sights, sounds, experiences- onto a Blue-ray to show your grandkids (who, by that time, will have something better than Blue-rays). Imagine being able to
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
- L14:35 strongly suppresses
- L17:118 excites comma after a clause ending
- L19:726 suppresses comma inside large numbers
- L18:366 excites comma inside numbers and dates
- L16:240 excites comma after 'sometimes' or 'the truth'
- L11:154 suppresses commas after citation numbers
- L9:77 excites commas in citations and name lists
- L4:3 excites commas after citation fields
- L17:585 excites
- L16:241 excites
- L18:93 suppresses
- L19:1036 suppresses
- L15:630 suppresses
- L10:335 excites commas in catalog, date, and list metadata
- L5:16 excites
- L12:1557 suppresses comma after an introductory phrase
- L4:96 excites
- L16:1059 excites
- L15:286 excites
- L19:1363 excites
- L3:71 suppresses
- L16:1939 excites line break mid-sentence in wrapped text
- L15:2227 weakly suppresses
- L15:1213 weakly suppresses
Feeds later layers
- L23:184 excites
- L25:880 excites comma after a cause or effect clause
- L22:144 excites
- L22:576 suppresses
- L21:344 suppresses
- L27:741 excites
- L23:839 excites comma before a trailing 'with' clause
- L27:291 suppresses
- L27:905 weakly suppresses
- L26:236 weakly suppresses
- L24:661 weakly excites
- L27:677 weakly excites
- L21:776 weakly suppresses
- L26:765 weakly suppresses
- L21:700 weakly excites comma before a new clause
- L26:586 weakly suppresses
- L24:572 weakly suppresses
- L27:86 weakly excites
- L21:113 weakly excites
- L23:1645 weakly suppresses
- L21:1604 weakly excites comma before a trailing phrase
- L25:306 weakly suppresses
- L25:468 weakly excites
- L25:216 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
- which
- 以及
- 也就是
- 尤其是在
- leading
- 包括
- 尤其是
- including
- especially
- 其中包括
Suppresses
- 、
- ),
- 的故事
- 的好处
- ,
- 的不同
- ,
- aven
- 的选择
- 的要求