the digit 2 inside citations
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. 4 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 reacts to the number 2 when it sits inside a code, reference or scripture number, such as a DOI, a Bible verse or "USB 2.0". It is quiet on the "2" in units like km2. Steering it did not produce 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.
-
LD, VAT GB238 5548 36<|endoftext|>Medical patients would probably be
-
94A21952571. https://dx.doi.org/10.2305/IUCN.UK.2016-2.RLTS.T6794A21
-
1 Jewish Holiday Calendar⏎|Biblical Name||Modern Holiday||Hebrew Date||English Date|⏎|Feast of Trumpets (Lev. 23:23–25; Num
-
circumstances over which they had no control. A quick check of Jesus’ parables shows the relationships between slaves/servants and wealthy people. (See Matthew 20:1-16, Luke 16:1-13
-
.⏎According to the Torah, in Exodus 12:16-20; Leviticus 23:5-8; Numbers 28:16-25
-
spend about a quarter of the moon's orbit protected by Earth's magnetosphere (Geophysical Research Letters, DOI: 10.1029/2007GL030507). "
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: no, not a lot. I don't know that the other... maybe not. Maybe not the time for me the space in. There is a lot. So, what? If I can't be, is there 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.
-
|⏎|• Total||1.09 km2 (0.42 sq mi)|⏎|• Density||13,000/km2 (34,000/sq mi)|⏎The Central Business District
-
83 sq mi (210 km2) 13%%|⏎|• Density||401/sq mi (155/km2)|⏎|Time zone||UTC−7 (Mountain)|⏎|• Summer
-
.6 square miles (250 km2) of it is land and 0.6 square miles (1.6 km2) of it (0.67%) is water.⏎As of the census
-
conditions. We compared weight variation and survival between control groups and mixed groups during three years of monitoring. We found both weight loss and high mortality in E. orbicularis of the mixed groups. This study argues for applying a precaution principle and stopping
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
- L3:279 suppresses
- L3:352 excites
- L3:2825 suppresses
- L0:418 excites
- L1:2536 weakly excites
- L1:82 weakly excites
- L3:631 weakly suppresses
- L0:17 weakly excites
- L1:892 weakly excites
- L0:949 weakly suppresses
- L0:1725 weakly excites
- L2:1936 weakly excites
- L3:2408 weakly suppresses
- L0:2877 weakly suppresses
- L3:986 weakly excites
- L3:2485 weakly suppresses
- L0:21 weakly suppresses last digit of years like 1970
- L2:174 weakly excites
- L3:861 weakly suppresses
- L3:1523 weakly suppresses
- L3:343 weakly excites
- L3:3031 weakly excites
- L2:2366 weakly suppresses
- L2:2935 weakly suppresses
Feeds later layers
- L17:1412 excites
- L16:1220 excites
- L17:128 excites
- L19:723 excites
- L24:1726 excites
- L18:122 excites
- L19:418 excites
- L20:1378 excites
- L9:1032 excites
- L9:13 excites
- L17:840 excites
- L19:851 excites comma inside numbers with thousands
- L18:2237 excites
- L19:258 excites
- L10:715 excites
- L22:801 excites the digit 1 in numbers
- L15:12 excites
- L15:1285 excites
- L26:1296 excites word-start 'mult' fragment
- L12:713 excites
- L10:50 excites
- L21:1863 excites
- L8:20 excites
- L17:157 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
- 请点击
- ��
- peace
- authorize
- beth
- 殓
- 瑁
- estate
- бит
- 响
Suppresses
- 符号
- букв
- 形式
- ides
- fractional
- "-"⏎
- symbols
- earer
- 超越
- jem