the zero completing 10 or 100
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. 9 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 the final '0' of a number like 10, 100 or 360 when the number is part of a calculation, such as percentages or decibels. The model's written output did not show a clear change in topic when it was turned up or down.
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.
-
case of a number, you can use the following formulas: 1350pc, x, or %. Then, multiply the result by 100 to get the percent. This method is also useful for finding the exact value of
-
I be MAX.⏎Express the PSNR in decibels. From Step 1, we have the decibel value LdB as LdB = 10 log10 (P1/P0). Now let P1 = MAX^
-
that are in the same units of measure. The decibel is 0.1 bel, so the decibel value LdB is LdB = 10 log10 (P1/P0).⏎Define the mean squared error (M
-
to know its total mark. In order to find this, use basic arithmetic. Divide the mark by the highest possible mark and multiply the answer by 100 to get the percent. Alternatively, you can try using
-
10 log10(MAX^2/MSE) = 10 log10(MAX/(MSE)^(1/2))^2 =
-
, 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 should be
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 a 4-year-old. I have a very good time with my friends and family. I love to play and explore, and I am really into the things I did. I am so much into the future. I don't
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
1) in physiology and medicine has evolved due to the accumulation of evidence supporting a major role of these proteins in the regulation of various biological and pathological phenomena.⏎Ubiquitin (2) is a 76-amino acid protein, which is
-
the same doctrine. And it’s the official teaching of the Catholic Church today – “In the most blessed sacrament of the Eucharist ‘the body and blood, together with the soul and divinity, of our Lord Jesus Christ and therefore
-
low levels of H3K27me3 that could predict prognosis.⏎"We could do it very simply, with a very
-
ubiquitin in basic cellular processes was cemented with the awarding of the 2004 Nobel Prize in Chemistry to ubiquitin research (3).Ubiquitin (Figure 1) can be attached to substrate proteins as a single moi
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
- L7:1690 excites
- L5:1069 weakly suppresses
- L8:1814 weakly suppresses
- L8:2401 weakly suppresses
- L5:417 weakly excites
- L8:2004 weakly excites
- L8:1723 weakly excites
- L2:2465 weakly suppresses
- L2:84 weakly excites
- L7:34 weakly suppresses
- L7:1526 weakly excites
- L8:2614 weakly excites
- L8:23 weakly excites
- L8:2129 weakly excites
- L8:1638 weakly suppresses
- L8:2331 weakly suppresses
- L8:1286 weakly excites
- L4:1015 weakly suppresses 'to' after 'not' (not to do)
- L8:2075 weakly excites
- L6:3039 weakly excites
- L7:2206 weakly suppresses
- L7:140 weakly excites
- L1:796 weakly suppresses
- L6:1902 weakly suppresses
Feeds later layers
- L10:1175 weakly excites
- L26:577 weakly excites
- L15:1481 weakly excites
- L12:1551 weakly excites
- L12:914 weakly excites
- L11:1738 weakly excites
- L25:1959 weakly excites
- L13:2098 weakly excites
- L11:2469 weakly excites
- L18:1245 weakly excites
- L12:2969 weakly excites
- L21:1009 weakly excites
- L23:249 weakly suppresses
- L16:655 weakly suppresses 'see' cross-references and web boilerplate
- L12:1911 weakly excites
- L12:2293 weakly excites
- L14:2449 weakly excites
- L14:1063 weakly excites
- L12:753 weakly excites
- L14:2431 weakly excites
- L23:2826 weakly suppresses
- L24:174 weakly excites
- L16:2713 weakly suppresses
- L12:556 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
- inters
- ippi
- $(
- $k
- 럼
- омер
- 然
- ows
- redits
- \xf
Suppresses
- озд
- azed
- Ashe
- щик
- occupy
- ysize
- redesigned
- pz
- ystack
- เสร