unclear: opening quote marks after a line break
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. 26 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 is hard to pin down. Its strongest reactions come on an opening quotation mark that starts a new paragraph in news-style text, but the steered outputs do not share a clear theme.
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.
-
apweed varieties also are major problems in the Four Corners area. Russian knapweed is the most common, overtaking fields and empty lots like wildfire.⏎"Russian knapweed spreads by black,
-
Coast — Did the Coast Change?" and calling on citizen volunteers to study aerial photographs taken before and after extreme storms to determine if or how the landscape changed.⏎"Computers cannot yet automatically identify damages and geomorphic changes to the coast from the
-
been directly observed and quantified in the thicker areas of the planet's shell, Calais said.⏎"Such dyking events had been included in theories, but researchers had never before been
-
to play outside," she says.⏎However, she stresses, indoor space is limited, and these homes had on average four people per room living in the homes.⏎"Outdoor play is
-
., who led the research. Her team is presenting the work this week at a scientific meeting, Purines 2010, in Barcelona, Spain.⏎"Acupuncture has been a mainstay of medical treatment in certain parts of the world
-
al collections manager at the Field Museum, Chicago, US, said hearing the news that the monkey belonged to a new genus "sent shivers down my spine".⏎"Simply put, the genetics said that it was closely related to baboons, but
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.
-
) законодательная власть⏎4) Associate Justice d) беспартийная система⏎5
-
,000 people every year.⏎Also known as vestibular schwannoma, acoustic neurinoma, or acoustic neurilemoma, an acoustic neuroma is a
-
ledge SJ, Niebuhr DW, Scher AI, Munger KL and Ascherio A⏎Multiple sclerosis (MS) is a chronic inflammatory demyelinating disease of the central nervous system of unknown etiology
-
s, flush residues after fossil fuels are burned and cool power plants.⏎Coal uses up to 3.2 cubic metres of water per megawatt hour (MWh) of electricity produced, gas uses up to 1.7 m3
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
- L19:43 excites
- L16:255 excites
- L2:1620 excites
- L16:3 suppresses
- L19:129 excites
- L23:1408 suppresses
- L5:97 suppresses
- L16:17 suppresses
- L3:395 suppresses
- L19:118 excites
- L19:89 excites
- L19:233 suppresses
- L4:345 excites
- L16:137 excites
- L20:344 excites hyphen inside compound words
- L16:7 suppresses
- L22:866 excites
- L22:271 suppresses
- L20:441 excites
- L2:1869 suppresses
- L24:248 excites
- L2:761 suppresses
- L2:404 excites
- L0:132 suppresses
Feeds later layers
- L27:248 suppresses
- L27:128 excites unclear: opening quotes and closing brackets
- L27:627 suppresses
- L27:1425 suppresses
- L27:172 suppresses
- L27:1007 suppresses
- L27:740 weakly suppresses
- L27:1520 weakly excites
- L27:2125 weakly excites
- L27:1019 weakly excites
- L27:2439 weakly suppresses slash inside web addresses
- L27:29 weakly suppresses
- L27:873 weakly suppresses
- L27:2229 weakly suppresses the letter h starting a word fragment
- L27:2889 weakly suppresses
- L27:657 weakly suppresses
- L27:1946 weakly suppresses the letter z inside words
- L27:981 weakly suppresses
- L27:929 weakly suppresses
- L27:541 weakly suppresses
- L27:2451 weakly excites
- L27:681 weakly suppresses
- L27:2518 weakly suppresses the letter f starting a word
- L27:1459 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
- vant
- .terminate
- ��
- Macron
- vably
- �
- [%
- mint
- seller
- 瘗
Suppresses
- ></
- ><
- > ⏎
- >{{$
- In
- It
- 知道
- >{
- The
- Well