zero in powers of ten
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 zero in scientific notation, like the 0 in 3 x 10^8 or 10^67. When it was turned up, the model's writing drifted toward powers of ten and big-number maths.
Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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.
-
than in air, water is said to have a greater optical density than air.⏎Substance Speed of light (m/sec)⏎Vacuum 3 x 108⏎Air 3 x 108⏎Water 2.2
-
when written as 52!, but it can also be written in scientific notation as 8.065817517 × 10^67. If I attempt to write it using normal numerals, it
-
0, and mℓ = 0 is where σ = r/a0 and a0 is the Bohr radius (5.29 × 10-11 m). Calculate the position of the nodes for this wave function. Write
-
In air, light travels at 3 x 108 m/sec but in water, it travels only about 2.258 x 108 m
-
oltzmann formula:⏎E = εσT4⏎where ε=emissivity, a material property; σ = 5.67×10-8 ; T = temperature in Kelvin (absolute
-
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, including
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 have been working on this. I have been working on it. I am ready to share it. I am ready to share this. I am ready to share this. Now, I will share it. I am ready to share this
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
DLR, and NASA/JPL.⏎More information about Dawn is online at http://dawn.jpl.nasa.gov.⏎Image credit: NASA/JPL-Caltech/UCLA/MPS/DLR/IDA<|endoftext|>How To Guide⏎How to do a
-
our Big Blue Oasis in space:⏎Click on the globe to see enlargements of the 1400 or so photographs NASA received.⏎Courtesy NASA/JPL/University of ArizonaAn National Geographic website article says the next rover mission to the planet
-
voting rights, unless such rights are specifically withheld, as in special classes of “nonvoting” shares. The common stockholders’ legal rights may also include preemptive rights to maintain
-
93nm, deep in the ultraviolet range, and have been instrumental in driving the semiconductor industry from 90nm geometries down to 28nm. Unfortunately, 193
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
- L9:1525 suppresses
- L9:2397 suppresses
- L8:2324 suppresses
- L9:5 excites
- L5:883 weakly excites
- L9:1740 weakly suppresses
- L9:2628 weakly excites
- L6:631 weakly excites
- L5:1608 weakly suppresses
- L8:2846 weakly excites
- L8:360 weakly suppresses
- L6:2689 weakly excites
- L3:1529 weakly suppresses
- L8:7 weakly excites
- L6:1112 weakly excites
- L8:16 weakly suppresses
- L7:1871 weakly excites
- L9:1916 weakly suppresses
- L0:1082 weakly suppresses
- L9:2976 weakly suppresses
- L1:1546 weakly suppresses
- L8:1533 weakly excites
- L8:1648 weakly suppresses
- L8:3 weakly excites
Feeds later layers
- L14:3032 excites
- L19:1242 excites
- L11:2380 excites
- L13:2285 excites
- L19:1542 excites
- L16:1248 weakly excites
- L23:2436 weakly excites
- L20:1668 weakly excites
- L19:2313 weakly excites
- L14:1875 weakly excites
- L23:2146 weakly suppresses
- L17:2479 weakly excites
- L12:2172 weakly excites
- L13:2915 weakly excites
- L18:2199 weakly suppresses
- L18:798 weakly excites
- L11:1106 weakly excites
- L11:2532 weakly excites
- L19:1365 weakly suppresses
- L21:2187 weakly excites
- L16:1325 weakly excites
- L25:895 weakly excites
- L16:1077 weakly excites
- L18:1029 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
- кос
- 很久
- 亿
- phosphate
- ione
- osaur
- Jahres
- erals
- Podesta
- ศา
Suppresses
- sắc
- دن
- ordon
- chatt
- 和地区
- onStop
- ="%
- roat
- ADDING
- 内容