zero inside a number
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. 10 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 '0' that sits inside a number, like 50, 40 or 1,036, in tables, statistics and measurements. Turning it up made the model write unrelated, rambling text.
Written by Claude from this neuron's activations and steering outputs. Confidence: low. 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.
-
||190 beats per minute|⏎|35 years||93-157 beats per minute||185 beats per minute|⏎|40 years||90-153 beats per minute||180 beats
-
(49.6% girls; mean age, 13.5 years), 582 in 10th grade (50.0% girls; mean age, 15.4 years), and 5
-
vised)|⏎|Displacement:||925 long tons (940 t; 1,036 short tons)|⏎|Length:||205 ft (6
-
5% CI, 1.7-1.9) for 2009 to 2012 to an RR of 3.0 (95% CI
-
McKinley, North America⏎At the Denali National Park in the US state of Alaska lies the 20,320 ft (6,194 m) Mount McKinley or Denali (its Koyukon
-
have walked on land. These include the mighty Argentinosaurus and Alamosaurus, thought to have tipped the scales at around 73 tonnes (80 tons) each. Despite not quite attaining those sizes, the new species stands
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: If I have a lot of time, and if I have a lot of time, if I have a lot of time, if I have a lot of time, if I have a lot of time, if I have a lot of time
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
meditation on the meaning of the story of Lot's wife.⏎This article is available to subscribers only. Please subscribe for full access—subscriptions begin at $2.95. Already have an online account? Log in now. Already a print subscriber
-
to three different classmates. (No fewer than 100 words). Please remember this post and your replies to classmates are due Thursday by 11:59 pm. You are required to cite material from your reading in your post.
-
;⏎P0 = 0x80;⏎P0 = 0x98;⏎P0 = 0xC0;⏎P0 = 0xff;⏎} while(1); // Endless loop<|endoftext|>Wisconsin Well Water:
-
:59 PM MDT · Views: 1,130⏎HUMAN E
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
- L9:1032 excites
- L5:2950 weakly excites
- L3:2928 weakly suppresses
- L4:201 weakly excites
- L9:2379 weakly excites
- L0:1400 weakly suppresses
- L4:2878 weakly suppresses
- L2:2043 weakly excites
- L9:316 weakly excites
- L1:1070 weakly suppresses
- L5:2359 weakly excites
- L7:780 weakly suppresses
- L7:2126 weakly suppresses
- L9:802 weakly excites
- L7:2423 weakly suppresses
- L9:1326 weakly suppresses
- L7:2837 weakly excites
- L9:1957 weakly excites
- L8:1086 weakly excites
- L8:2027 weakly suppresses
- L1:1862 weakly excites
- L3:2426 weakly suppresses
- L6:1454 weakly excites
- L7:2448 weakly excites
Feeds later layers
- L11:560 suppresses
- L11:689 suppresses
- L14:2654 weakly excites
- L14:2834 weakly suppresses
- L19:2311 weakly excites
- L23:1406 weakly excites
- L20:2395 weakly excites
- L20:2611 weakly excites
- L23:920 weakly suppresses
- L16:3070 weakly excites
- L26:1259 weakly suppresses
- L26:2032 weakly excites 'to' or 'from' in 'from one X to another'
- L25:3059 weakly suppresses
- L19:843 weakly excites
- L23:1049 weakly excites
- L20:3007 weakly excites
- L19:1679 weakly suppresses
- L23:1648 weakly suppresses
- L21:2816 weakly suppresses
- L16:2579 weakly suppresses
- L20:2665 weakly excites
- L18:2399 weakly excites
- L25:981 weakly suppresses
- L11:553 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
- yw
- 我希望
- grap
- newY
- ACHER
- ไหว
- wished
- /world
- brig
- pneum
Suppresses
- asename
- 不确定性
- istics
- ijken
- Keyword
- ằng
- "⏎⏎⏎
- 存量
- 异议
- ilogue