the digit 1 in 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. 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
This neuron responds to the number 1 (and sometimes zeros) inside numbers, formulas, and maths problems. Turning it up made the model's text wander into unrelated topics rather than numbers.
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.
-
6 inches. This radium emits tiny amounts of radon into the atmosphere. On average, one part of radon is present in 1 x 1021 part of air.Radon is also present in some spring waters,
-
areas of mainland China. In 1955, 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 Renmin
-
energy is expected to be saved over ten years when using a solar water heater compared to the electricity consumption of an equivalent electric water heater. 1 x STC is 1 x MWh of electricity saved over 10 years = 1
-
know how to solve that. We can multiply by 5 and divide by 5.⏎- 1/5 times 5 is just going to be 1. So we haven't changed the value of the expression.⏎- But when we
-
of all, it is important to clarify what exactly a percentage is, and this is nothing more than a quantity or number that represents a fraction of the number 100.⏎In other words, it is a part of a general number that
-
as the true power of the circuit.⏎(i) For a purely resistive circuit, the phase angle between voltage and current is zero and cos ϕ = 1.⏎∴ Pav = VRMS IRMS⏎(ii) For a purely inductive
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 it's possible to get a 12.5. Let me check... Let me get it right... Okay, 12.5, 12.5. Let me get it right... Okay, 1
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
and attractive indoor environments. It is no surprise that this sustainable building boom has also led to the construction of more efficient schools. Studies show that net-zero K-12 schools consume
-
or uncertain? "Fluctuations in labor input in a RBC model, not caused by real wage fluctuations, should be negatively correlated with consumption."⏎Choose 1 out of the following 2. Provide a strict formal argument. Worth 2
-
10 log10(MAX^2/MSE) = 10 log10(MAX/(MSE)^(1/2))^2 =
-
. As the first black woman we have ever seen in the Tunisian Parliament, I am proud to be part of that minority, always.”⏎Ks<|endoftext|>Select one (1) column from a dataset with whole non-negative integers. A good dataset
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
- L10:848 suppresses
- L9:1373 suppresses
- L10:1284 excites
- L10:1602 excites the number 1 as a unit value
- L6:902 excites
- L4:961 excites
- L7:753 excites
- L5:559 suppresses
- L10:2192 suppresses
- L3:244 suppresses
- L9:977 suppresses the last digit of a decade year
- L8:972 excites
- L10:773 weakly excites
- L2:1225 weakly suppresses
- L9:2714 weakly excites
- L6:18 weakly excites
- L10:722 weakly suppresses
- L10:548 weakly excites zeros inside large numbers
- L7:39 weakly excites
- L1:2025 weakly excites
- L5:920 weakly excites
- L7:2305 weakly suppresses
- L5:91 weakly excites
- L2:1462 weakly suppresses "factors" after "other" or "several"
Feeds later layers
- L13:791 excites
- L13:667 excites zeros inside decimal numbers
- L12:641 suppresses
- L16:950 excites
- L16:1012 excites 'one' before 'of' in 'is one of the'
- L18:2221 excites
- L13:919 excites
- L13:2411 excites
- L15:2061 excites
- L18:726 suppresses
- L14:1170 excites
- L13:2568 suppresses
- L13:564 suppresses
- L14:1414 excites
- L12:2435 weakly excites
- L12:370 weakly suppresses
- L15:2651 weakly excites
- L18:2793 weakly excites
- L13:2485 weakly excites
- L12:1615 weakly excites
- L16:1420 weakly suppresses
- L14:2062 weakly excites
- L16:2576 weakly suppresses
- L16:1856 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
- ESIS
- .setCharacter
- ease
- .CompareTo
- ['',
- զ
- üğü
- 日起
- devez
- WITHOUT
Suppresses
- ajo
- intr
- ecure
- repid
- ensive
- ressive
- Peninsula
- ench
- ids
- パー�