the letter d inside words
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. 27 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 lights up on the letter d when it starts a word piece, like the d in dim, dementia, ditches or dimes. When it was turned up, the model's replies got stuck repeating filler words such as 'about about about', which does not clearly relate to the d pattern.
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.
-
ME.⏎DIM is traced to Old Norse dimmr (dark) andIEbasedhem (smoky). Russian smoke is dim; in⏎Turkish it is duman.⏎Edenic covers the silence and senselessness of DUMB. דמי
-
of all the different types of dementia (including younger-onset dementias), from the most-recognized - Alzheimer's - to the less-f
-
high on inside fence, and well-preserved section of ditch c 3m below rampart. On the SW and SE sides the defences consist of twin ditches with a medial rampart. The ditches are slight and on the SW
-
Accusative Nominative Accusative ichmich dudich erihnsiees wiruns ihreuchsie/Sie⏎Pre
-
όφῳ in Einklang zu bringen, ist es unumgänglich nothwendig dass der Ausdruck Roma quadrata technisch in doppeltem Sinne gebraucht wurde; einmal zur Be
-
700 pounds. Mechanical fastening is difficult because of a tendency for splitting. Predrilling for nails and screws is advisable.⏎Drying. The wood dries slowly with some risk of checking, end splitting, and warping. Slow
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:ismismismismismismismismismismismismismismismismismismismismismismismismismismismismismismismismismismismismismismismismismismismismismismismism
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
sons. He loved them very much. They always used to help one another<|endoftext|>Ramses III, Ramses also spelled Ramesses or Rameses (died 1156 bce, Thebes, Egypt), king of
-
trigger.” 26 U.S.C. § 5845(b). The machine gun was first widely used during World War I, where it “demonstrated its murderously effective firepower over and over again.” William Rosenau
-
with a specialist, rural emergency room physicians are more likely to adjust diagnosis and course of treatment for injured children, research shows.⏎The use of technology to link far-distant practitioners
-
combination of employer and employee contributions and offer a range of options, including Health Maintenance Organizations (HMOs), Preferred Provider Organizations (PPOs), and High-Deductible Health Plans (HDHPs
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
- L23:400 excites
- L5:196 suppresses the 'D' in vitamin D
- L10:180 excites
- L17:300 suppresses
- L0:233 excites
- L19:372 suppresses
- L22:1511 suppresses
- L2:110 excites
- L20:200 excites
- L20:1659 excites
- L1:10 excites
- L21:1912 weakly excites
- L18:105 weakly suppresses
- L1:773 weakly excites
- L23:206 weakly suppresses
- L22:473 weakly suppresses
- L26:1564 weakly excites
- L18:704 weakly suppresses
- L18:554 weakly suppresses
- L4:1828 weakly suppresses
- L19:441 weakly excites
- L18:569 weakly excites
- L16:461 weakly suppresses
- L19:526 weakly suppresses
Feeds later layers
None: this is the last layer.
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
- دد
- daf
- dad
- gable
- ться
- 和发展
- Collection
- mega
- 也应该
- guna
Suppresses
- ism
- ifer
- iss
- ove
- edic
- isting
- immers
- icism
- edb
- 记者了解