blank slot before a date
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 the empty spot right before a day or date, such as "on [blank] 24 April" or "on [blank] 26 September 1844". Turned up, the model's text changed but did not clearly show a date 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.
-
, Manchester and finishing at the Manchester Museum on Oxford Road.⏎The Manchester Museum will also be hosting a special 'A History of the World Big Saturday' on Saturday 24 April.<|endoftext|>Last month, President Biden issued an executive order directing government
-
the most⏎controversial clergyman in the history of the Victorian Presbyterian Church,⏎Strong was born at Dailly, Ayrshire, Scotland on 26 September 1844 the son⏎of Rev. David Strong
-
ending the First and Second World wars. A statue of him was erected to commemorate him in London's Parliament Square.⏎Early life and education⏎He was born on 24 May 1870, at the family farm, Boven
-
unaccounted for.⏎The mission was extremely hazardous for the members of the Simcol detachments, particularly following Hitler's infamous ‘Commando Order’ issued on 18 October 1942, ordering the annihilation of all men
-
) which⏎received the assent of the President on the 15th April 1962 and was published in the⏎Official Gazette on the 2nd May 1962. The object was equitable distribution of land
-
for the boys and they’re loving the 21 Century learning. We hope to see you at our World Travel Expo between 2-3pm on Wednesday 8 June.⏎by Lee White, Junior Boys Year
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: how to place, as a natural, in the future, the necessary and necessary and necessary and needed, and then to place something, as a natural, in the next, as a natural, and so on, as a natural, in
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
to come.⏎Announcement: We are moving to a new web site. There may be some disruption of service from 3/23/13-3/24/13<|endoftext|>…In 197
-
extremes in temperatures, sunlight, tight-fitting clothing, etc.)<|endoftext|>ERIC Number: ED348648⏎Record Type: Non-Journal⏎Publication Date: 1992-Apr⏎Reference Count: N/A⏎Finding
-
area.⏎Neil burn Drive was created.⏎Canmore place was erected.<|endoftext|>ERIC Number: ED379961⏎Record Type: RIE⏎Publication Date: 19
-
at 361-396-2831. Or if you need to make a claim with Flint Hills you can do so at 1-800-<|endoftext|>This is good book for TPR Excellent book that
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
- L16:2810 suppresses
- L17:1872 excites
- L19:2786 excites
- L0:2685 suppresses
- L0:1075 weakly excites
- L19:2473 weakly excites
- L18:476 weakly suppresses
- L13:1545 weakly suppresses
- L19:2444 weakly excites
- L18:560 weakly suppresses
- L19:1333 weakly suppresses
- L16:899 weakly suppresses
- L2:1502 weakly excites
- L17:1246 weakly excites
- L19:2242 weakly suppresses
- L1:179 weakly suppresses
- L7:1034 weakly suppresses
- L11:223 weakly suppresses
- L2:94 weakly suppresses
- L15:855 weakly excites
- L16:2908 weakly suppresses
- L11:1484 weakly excites
- L1:2546 weakly excites
- L6:605 weakly suppresses
Feeds later layers
- L24:600 strongly excites
- L23:176 strongly suppresses
- L24:401 strongly suppresses 'it' referring back to a noun
- L24:1080 excites
- L24:120 suppresses
- L25:586 suppresses
- L23:625 excites
- L27:650 excites
- L21:1108 suppresses
- L25:2055 excites
- L23:1210 excites
- L24:241 excites
- L26:72 excites 'they' or 'we' after where or who
- L21:275 suppresses
- L21:468 suppresses
- L26:2947 excites
- L22:448 excites
- L23:1071 excites
- L26:874 excites
- L23:287 excites 'are' after a plural subject
- L23:564 excites non-English text, mid-phrase
- L24:314 suppresses
- L23:311 excites
- L27:577 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
- 都
- 都有
- _are
- are
- 都很
- 都可以
- 都没
- 都没有
- 都不
- 都非常
Suppresses
- blr
- AILS
- manages
- astes
- heiro
- recognizes
- ости
- Handles
- needs
- speaks