commas before an elaboration
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 or a word that opens a list or explanation, such as the comma after 'There are many ways to set up multi-factor authentication'. The peaks are modest. Turning it up made the model repeat the word 'and' over and over.
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.
-
221BC). Unlike most Chinese idioms, 毋忘在莒 strangely has actually two stories of origin, which gave this id
-
. A seizure is a short episode of symptoms caused by a burst of abnormal electrical activity in your brain. Different parts of the brain control different functions of your body, so the symptoms that occur during a seizure will depend on where the abnormal burst of
-
a copy of the report go to the NOAA news website www.noaanews.noaa.gov⏎– Erich Luening⏎Print this page<|endoftext|>Today is Sunday, Feb. 3, the 34th day of 201
-
chede and colleagues are optimistic that it could become an integral part of the semiconductor industry. Lithography is widely used in the industry to make two-dimensional circuit boards, but a convenient method for creating three-dimensional<|endoftext|>Embedding a Racial Equity Perspective
-
attach themselves to shell fragments so they have something to latch on to as they begin to grow their own shells. The farmer said any type of shell fragments will do, they don’t have to be oyster shells.⏎Once the young
-
mobile app, web app, or website as part of your business offering, multi-factor authentication is a must. There are many ways to set up multi-factor authentication, but some popular, secure methods include:⏎- Email – let the user receive a
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: with the previous, with the previous, with the previous, with the previous, with the previous, with the previous, with the previous, with the previous, with the previous, with the previous, with the previous, with the previous,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
��性病的⏎anachronism /əˈnækrəˌn
-
the data.⏎D. Types of Splits⏎There are two types of splits in decision trees: continuous and categorical. Continuous splits are based on a threshold value, such as the split in a regression task where all instances below the threshold value are
-
weather’ and ‘climate.’ Weather represents short-term atmospheric conditions in a specific area, while climate is the average of these conditions over extended periods. This distinction is key
-
use it to access patient records and share test results. Governments use it for elections, services and disaster alerts. The global financial system—bank transfers, stock trades—increasingly relies on it.⏎Because of abstraction and the distributed nature of cloud
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
- L19:39 excites
- L2:422 excites
- L19:43 excites
- L16:36 suppresses
- L4:490 suppresses decade digit in years like 1930s
- L5:97 suppresses
- L18:569 excites
- L16:90 suppresses
- L9:77 excites commas in citations and name lists
- L0:579 suppresses
- L0:132 suppresses
- L16:177 suppresses zero digit inside dates and codes
- L2:202 suppresses "of" after "many" or "which"
- L16:7 suppresses
- L4:489 excites
- L19:118 excites
- L6:840 suppresses
- L5:1278 suppresses
- L17:215 weakly excites
- L0:304 weakly excites
- L19:1221 weakly suppresses
- L18:366 weakly excites comma inside numbers and dates
- L17:213 weakly excites
- L16:34 weakly excites
Feeds later layers
- L24:205 strongly excites
- L26:347 strongly excites the 're' in self-reliant and similar
- L27:289 strongly excites
- L26:65 strongly suppresses
- L22:791 strongly excites
- L23:379 strongly excites
- L27:172 excites
- L25:6 suppresses
- L21:739 excites 'both X and Y' pairing words
- L22:632 excites
- L25:1270 suppresses
- L27:1902 weakly suppresses
- L25:583 weakly excites
- L21:2699 weakly suppresses
- L26:434 weakly suppresses
- L22:286 weakly excites
- L27:1593 weakly suppresses
- L27:1121 weakly suppresses
- L24:921 weakly suppresses
- L25:120 weakly suppresses
- L22:2060 weakly excites
- L26:859 weakly excites
- L25:1786 weakly suppresses
- L26:1150 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
- 和
- and
- 和
- 和个人
- và
- and
- 和地区
- 和技术
- 性和
- и
Suppresses
- /or
- ари
- ape
- นา
- elif
- istica
- rade
- obi
- erta
- ᴇ