comma before a qualifying clause
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. 19 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 commas in the middle of a sentence, right after a full statement, where the writer is about to add a qualifier. When it is pushed down, the model tends to write the word 'especially' over and over, so it seems to hold that word back.
Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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.
-
are the serosa, muscularis externa, submucosa and mucosa. This layered arrangement follows the same general structure in all regions of the stomach, and throughout the entire gastrointestinal tract. The outer layer of the stomach wall is smooth
-
transfer protocol (HTTP) server that runs in the background as a daemon. Because⏎it is a daemon, it must do all the things that background applications do, namely,⏎read its configuration files, go into the background, shut down when told
-
with a spread of about 50 feet. The compound leaves give it a lacy, airy appearance. This tree will tolerate a wide range of soil conditions, but it does require good drainage. Leaves will turn orange to red, and sometimes
-
as long as it did or we would have suffered a gap.”⏎But Landsat 6 was developed and launched outside of NASA’s control by a provate company, and Irons
-
south pole, providing further evidence of precipitation.⏎Elizabeth Turtle of John Hopkins University in Laurel, Maryland, and colleagues compared images from two flybys of the moon, taken on 3 July 2004 and 6 June 2
-
useful, but what if you have to make several copies, or you just want to make a complete backup of a hard drive? Drive images let you put everything, and we mean everything, from your hard drive in one big file.⏎With an
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: especially, especially, especially, especially, especially, especially, especially, especially, especially, especially, especially, especially, especially, especially, especially, especially, especially, especially, especially, especially, especially, especially, especially, especially,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
a few of the many ways to use edtech:⏎1. Conquering Teacher Shortage and Limited Course Offerings.⏎Finding qualified teachers can be a challenge, especially at this point in the year, but that doesn’t have to stop you
-
further costs.⏎4. Challenges and Considerations:⏎a. Affordability: Medical insurance premiums can be costly, especially for those without employer-sponsored coverage. Government subsidies and tax credits can help make
-
inflamed, it can swell and fill with pus. This can cause severe pain, nausea, and vomiting. Swelling is therefore an important symptom to watch for, especially if it is accompanied by other symptoms of appendicit
-
ism. Some of the most famous ukiyo-e artists include Katsushika Hokusai,<|endoftext|>Weight loss can be a major struggle for some people, so much so that they turn to gastric bypass surgery to help them achieve their weight
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
- L17:118 excites comma after a clause ending
- L14:35 suppresses
- L16:240 excites comma after 'sometimes' or 'the truth'
- L17:2004 excites
- L18:1523 excites
- L18:366 excites comma inside numbers and dates
- L18:2376 suppresses
- L18:569 excites
- L18:100 excites
- L9:77 excites commas in citations and name lists
- L11:154 suppresses commas after citation numbers
- L17:2836 weakly excites
- L3:71 weakly suppresses
- L10:612 weakly excites full stop ending boilerplate or a title
- L16:241 weakly excites
- L0:178 weakly suppresses
- L0:1224 weakly suppresses
- L16:1748 weakly suppresses
- L5:2309 weakly excites
- L16:431 weakly excites
- L0:1301 weakly suppresses
- L2:483 weakly suppresses
- L18:1869 weakly excites
- L14:2254 weakly suppresses
Feeds later layers
- L20:1519 strongly suppresses
- L26:2236 strongly suppresses comma before an 'especially' aside
- L27:2047 strongly suppresses
- L23:720 suppresses
- L22:368 suppresses
- L22:158 excites
- L24:203 excites
- L26:236 suppresses
- L27:358 suppresses
- L27:213 excites
- L20:2082 suppresses
- L26:113 excites
- L20:622 excites
- L22:940 suppresses
- L27:747 excites
- L27:86 excites
- L21:436 excites
- L24:216 suppresses
- L26:22 excites the 'res' fragment inside words
- L26:606 weakly excites
- L20:882 weakly excites
- L26:80 weakly excites
- L26:388 weakly excites
- L21:384 weakly excites
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
- =
- {|
- iled
- =⏎⏎
- تها
- = ⏎
- {.
- =⏎
- |
- �
Suppresses
- especially
- 特别是
- 尤其是
- especially
- 尤其
- 尤其是在
- particularly
- 特别是在
- particularly
- Especially