digits inside numbers
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 lights up on the digits inside numbers, especially the zeros and last digits that finish a figure like 10,000 or $10. When it was turned up, the written text drifted into unrelated topics, so the steering did not show a clear effect.
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.
-
of the year using altars of various geometric configurations. India is largely credited with inventing the idea of zero and for developing place notation (ones, tens, hundreds, etc.), the decimal system, and the numeral system used today.⏎Already established
-
1 seconds). As a comparison, today’s hard drives take a few nanoseconds to record 1 bit of data. A nanosecond is 1000 picoseconds, making it very clear how much faster the heat pulse solution will
-
names organisms based on studies of their different characteristics. Helps to study the relationships between living and extinct species.⏎Why is taxomony important to scientist?⏎Domain, Kingdom, Phylum, Class, Order, Family, Genus, Species⏎List
-
55, a re-evaluation took place and a new Yuan Renminbi was introduced at a rate of 1 new Yuan to 10,000 old Yuan.⏎The Renminbi in Foreign Exchange⏎During the command economy
-
that may be linked to cancer (nonmuscle myosin II). In addition, inherited familial cardiomyopathies affecting 1 out of 500 people result from missense mutations in various cardiac muscle proteins, with mutations in beta
-
Here the children are asked to look at the grid of numbers and identify the missing numbers and then fill in the missing numbers to count from 1 to 25. In these missing number worksheets some of the numbers are already filled in to help
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: the one that is... if you look at it. The first thing that comes to mind is that, for example, and then, and then, and so on. The problem, the first thing, and the second, and the third
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
and don’t cut back on food and drink, we need to do more.⏎The American College of Sports Medicine agrees adults should get at least 150 minutes of exercise a week, though it explains this might be 20 to
-
at the time of a maltreatment investigation associated with the likelihood of remaining in foster care? NSCAW is a national longitudinal study of 5,501 children, aged 14 years or younger, who had contact with the child
-
July 10, the fire had damaged or destroyed nearly 130 buildings, injured over 20 people, and cut the lives of 19 Granite Mountain Hotshots woefully short.⏎Sparked by Lightning, Fueled
-
against similar survey data from 2009 to 2012.⏎Study participants included 570 students in 8th grade (49.6% girls; mean age, 13.5 years),
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
- L8:41 suppresses
- L9:657 suppresses
- L7:895 suppresses
- L0:208 suppresses the digit zero inside codes or numbers
- L1:1993 weakly excites
- L2:241 weakly excites
- L5:364 weakly suppresses
- L3:2824 weakly suppresses
- L4:1926 weakly excites
- L9:2379 weakly excites
- L9:3037 weakly excites
- L4:1343 weakly excites
- L6:567 weakly suppresses technical terms mid-phrase, mixed topics
- L4:394 weakly excites
- L8:1706 weakly suppresses
- L3:279 weakly excites
- L6:2354 weakly excites
- L3:230 weakly excites
- L1:2776 weakly suppresses
- L9:1926 weakly suppresses
- L7:1753 weakly suppresses
- L0:1788 weakly excites
- L5:424 weakly suppresses
- L4:667 weakly excites
Feeds later layers
- L13:2568 excites
- L12:946 weakly excites
- L15:2117 weakly excites
- L17:1522 weakly suppresses
- L12:1194 weakly excites
- L15:2768 weakly suppresses
- L18:1125 weakly excites
- L11:2080 weakly suppresses
- L11:1106 weakly excites
- L17:1795 weakly suppresses 'and' before a second verb phrase
- L21:734 weakly suppresses
- L23:615 weakly excites
- L27:1131 weakly excites
- L17:1580 weakly excites
- L12:2550 weakly suppresses
- L13:1019 weakly excites
- L24:3048 weakly suppresses
- L14:263 weakly excites
- L15:1066 weakly suppresses
- L17:3065 weakly excites
- L21:1244 weakly suppresses
- L13:2869 weakly suppresses
- L23:2121 weakly excites
- L12:1272 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
- 化
- brace
- strar
- 化的
- protected
- pu
- grid
- Constraint
- constraint
- gr
Suppresses
- 汲
- resident
- YLE
- ALLE
- alley
- oman
- enames
- 蔬
- ALLED
- poop