words right before 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. 0 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 small words like 'nearly', 'about', 'in' and 'align in' when a number is about to follow, such as 'nearly 400,000' or 'in 1995'. It is basically a 'a number comes next' detector. Turning it up did not bring out number-related text, so the steering samples do not confirm it.
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.
-
over.⏎The recent eruptions displaced nearly 400,000 people, according to the Disaster Management Agency.
-
infections.⏎Sub-Saharan Africa remains most severely affected, with nearly 1 in every 25 adults living with HIV and accounting for nearly 7
-
to nearly 459.67 degrees Fahrenheit below zero. “It works like optical
-
are aligned tonight?⏎- What planets will align in 2020?⏎- Which two planets have no moons at all?⏎-
-
Wrap about 18 inches of floss around your middle fingers, allowing about two inches between
-
action programs.⏎A crucial turning point came in 1995, when the University of California’s Board of Regents made the
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 am thinking about something. I am thinking about something. I am thinking about something. I am thinking about something. I am thinking about something. I am thinking about something. I am thinking about something. I am thinking about something.
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
entomologist and director of ASU's International Institute for Species Exploration.⏎Shocking pink dragon millipede (Desmoxytes purpurosea) from Thailand. According to its discoverers, this millipede's g
-
conference with an improved simulation, including better audience controls and much higher expectations. This time, the participants were told to steer a submarine through a 3-D undersea adventure. They were to capture the eggs of a sea monster. It was the
-
.⏎Here are some ideas!⏎A deep prehistoric jungle filled with dinosaurs⏎Ocean cliffs with Pterosaurs nesting, feeding, and filling the sky⏎An undersea landscape filled with prehistoric sea life⏎An unground maze of burrows and
-
sis.⏎4. Identify the drug, classify the type of reaction,<|endoftext|>History can teach us what will happen.⏎The modern State of Israel is the third commonwealth to be established on God’s Holy Land. The first was when Joshua conquered Can
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
None: this is the first layer.
Feeds later layers
- L18:377 excites number right after 'about' or 'less than'
- L19:1061 weakly excites
- L8:30 weakly excites blank slot right before a number
- L4:53 weakly excites blank slot before a number
- L1:764 weakly suppresses
- L22:824 weakly excites blank before an expected number
- L17:285 weakly excites
- L19:2521 weakly excites
- L25:242 weakly excites
- L2:1842 weakly suppresses
- L11:3018 weakly excites
- L9:519 weakly excites
- L10:220 weakly excites hyphen inside compound words
- L23:891 weakly suppresses
- L27:2506 weakly excites
- L5:1360 weakly suppresses
- L3:449 weakly suppresses
- L16:1752 weakly suppresses
- L2:1534 weakly suppresses
- L10:1242 weakly suppresses
- L3:21 weakly suppresses
- L3:107 weakly suppresses line breaks mid-sentence
- L27:1895 weakly excites
- L2:27 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
- <=$
- \data
- epy
- openhagen
- ató
- =size
- <data
- =data
- ONDON
- *time
Suppresses
- 人的
- -
- Horizontal
- 忽略
- ot
- Horizontal
- wo
- MLP
- 回事
- lt