comma after 'So' or intro phrase
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. 21 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 right after an opening phrase such as 'So,' at the start of a new sentence, as in 'So, what exactly does...'. Turning it down made the model produce strings of commas and symbols, and turning it up gave loose number lists.
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.
-
where they have pinpointed the differences between boats and canoes kayaks. But they should’ve been doing was pinpointing the technique differences and explained why. So, I will do exactly what needs to be done.⏎To understand the difference of techniques
-
for anyone with mobility issues, these car journeys are necessary; therefore, the analysis accounted for the population of adults who reported mobility issues in the NTS).⏎So, car trips for 1 mile or under add up to 1.8 billion
-
, making it difficult to distinguish.⏎Perhaps you have found a less common variety that doesn’t match your preconceived notion of what obsidian looks like. So, what exactly does obsidian look like?⏎Obs
-
marker. Battleship U.S.S. Ohio (approx<|endoftext|>“Snow White,” the beloved, first, full-length animated feature film, created by the Walt Disney Studios, is celebrating its 75th anniversary this
-
and smaller space.⏎In almost all eukaryotes, histone-based chromatin is the standard, yet in bacteria, there are no histones. So, how and why have histones become so entrenched in our chromatin structures during evolution
-
titanosaur have been revealed by researchers who excavated the bones from a site in central Russia. The first titanosaur ever discovered in the country, it shows unique features never before seen in other sauropod specimens.⏎The bones
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.
-
do everything in this world?⏎Close MATLAB with keyboard⏎Find the treasures in MATLAB Central and discover how the community can help you!⏎Choose a web site to get translated content where available and see local events and offers.
-
⏎Polygon Interior Angle Sum⏎Bit to Gigabyte⏎Find the treasures in MATLAB Central and discover how the community can help you!⏎Choose a web site to get translated content where available and
-
2-year-old to respect others. Begin with the basics (saying "please" and "thank you"), then try these six other strategies.<|endoftext|>This archived Web page remains online for reference, research or recordkeeping purposes. This page will
-
5 direct from Amazon Amazon page|⏎Setup an account with your affiliations in order to access resources via your University's proxy server⏎Configure custom proxy (use this if your affiliation does not provide a proxy)⏎|Through your library|⏎References found in
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
- L20:863 excites
- L20:1298 suppresses
- L19:726 suppresses comma inside large numbers
- L9:77 excites commas in citations and name lists
- L18:48 suppresses
- L15:123 suppresses
- L14:35 suppresses
- L16:240 excites comma after 'sometimes' or 'the truth'
- L14:51 excites colon that introduces a statement
- L19:380 excites
- L5:5 suppresses digit after a label like paper or type
- L17:62 suppresses
- L9:7 excites
- L14:348 weakly suppresses
- L20:1026 weakly excites
- L15:44 weakly suppresses
- L17:38 weakly suppresses
- L16:1596 weakly suppresses
- L0:68 weakly suppresses
- L11:18 weakly excites
- L6:228 weakly excites comma after a filler word
- L20:8 weakly suppresses
- L20:128 weakly suppresses
Feeds later layers
- L27:63 excites closing parenthesis in data lists
- L27:139 excites
- L27:433 suppresses
- L27:954 suppresses
- L26:153 suppresses
- L26:193 suppresses
- L27:1332 suppresses
- L27:1749 suppresses hyphen joining word parts
- L22:215 suppresses
- L27:77 excites
- L23:184 excites
- L26:22 excites the 'res' fragment inside words
- L27:25 excites
- L26:80 excites
- L27:838 suppresses
- L25:112 suppresses
- L27:71 excites
- L27:143 excites
- L27:327 weakly suppresses
- L22:428 weakly suppresses
- L27:191 weakly suppresses
- L26:94 weakly excites
- L22:16 weakly excites comma in dates and birth lines
- L26:345 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
- whereas
- 물을
- while
- ras
- Wikimedia
- 물
- {}".
- puis
- 何况
- #{@
Suppresses
- strument
- IRONMENT
- atomy
- alyzed
- ned
- adays
- alyze
- ishly
- icient
- xiety