colons in clock times
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. 11 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
The neuron responds to the colon in times of day, like '02:16 pm' or '17:47 GMT', in dates and timestamps. On the other side it dislikes certain religious and political names such as 'God' or 'Allah'. Turning it up produced unrelated text, so steering shows no clear 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.
-
the north although it could be overhead or elsewhere. For your best chance of sighting an aurora, try to look during the hours around local midnight (22:00-02:
-
I guess, it's not the first time it's happened...<|endoftext|>Last updated on September 17th, 2022 at 02:16 pm⏎Synodontis catfish are one of the most popular species
-
subject is much smaller than a human. Essential oils are highly concentrated and should<|endoftext|>Monday, October 5, 1998 Published at 17:47 GMT
-
SeaWeb/FIS)⏎Seafood producers tackle ocean acidification at annual Summit⏎Monday, February 21, 2011, 23:50 (GMT + 9)⏎From 31 January to 2
-
11:07:11 EDT⏎- How Martian winds make rocks walkThu, 8 Jan 2009, 14:01:08 EST⏎- MIT moves toward greener chemistryFri,
-
'Lose my life, my leg or my hand?': Modern-day slavery in India⏎10 July 2014 Last updated at 18:48 BST⏎It is a form of modern day slavery otherwise known as forced
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: "I think the United States is the first, the second, the third, and so on, the highest, highest, highest, and so on, the highest, highest, highest, and so on, the highest, highest, highest
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
of the nation" (White Paper on Tibet, Sep 1992). Chen Kuiyuan, a Chinese cadre appointed as leader of the Chinese Communist Party in Tibet in March 1992, has called on "in
-
107 of the Copyright Act 1976.⏎Fair son, the first thing I would teach thee is to set thine heart to love God; for unless he love God none can be saved.”⏎From Louis’ last instructions
-
is necessary for all new converts to recite this upon conversion to the new faith.⏎The first part of this declaration is that there is none worthy of worship except Allah. As we have previously mentioned, Allah is the personal name for the One God
-
the establishment of the People's Republic of Albania in January 1946, Albania became a rigid police state, dominated completely by the communist party and by Marxism/Leninism. Although Albania operated under the facade of constitutional rule, the
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
- L4:42 suppresses
- L3:1262 weakly excites
- L10:2802 weakly suppresses
- L10:495 weakly excites
- L8:12 weakly excites
- L3:136 weakly excites
- L4:2224 weakly excites
- L9:824 weakly suppresses
- L9:1480 weakly suppresses
- L8:4 weakly excites
- L3:2269 weakly suppresses
- L8:2404 weakly suppresses
- L10:1921 weakly suppresses
- L4:1259 weakly suppresses
- L9:2160 weakly excites
- L2:283 weakly excites
- L8:3020 weakly excites
- L7:1826 weakly suppresses
- L8:2930 weakly excites
- L10:2014 weakly excites
- L6:1247 weakly excites
- L7:1962 weakly suppresses
- L5:758 weakly suppresses
- L9:2955 weakly suppresses
Feeds later layers
- L13:657 suppresses
- L14:1681 excites
- L16:2360 weakly suppresses
- L22:1726 weakly excites
- L14:1153 weakly suppresses
- L14:2632 weakly excites
- L17:1591 weakly suppresses
- L22:1518 weakly excites
- L15:2056 weakly suppresses
- L15:1556 weakly suppresses
- L21:1133 weakly suppresses
- L24:2309 weakly suppresses
- L23:1082 weakly excites
- L14:540 weakly suppresses
- L24:909 weakly suppresses
- L15:281 weakly suppresses
- L25:2731 weakly suppresses
- L14:2106 weakly excites
- L13:431 weakly suppresses
- L15:509 weakly suppresses
- L24:1346 weakly excites
- L13:1848 weakly excites
- L20:2036 weakly excites
- L23:356 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
- 展品
- extras
- út
- ojis
- habitat
- pró
- ерь
- erte
- ophile
- statt
Suppresses
- acknowledgement
- própria
- 葭
- yourself
- commercial
- onest
- 鸱
- otten
- 爱你
- orton