the digit 5 in decimals like 1.25
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. 2 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 number 5 when it sits inside a decimal or a number, such as the 5 in 1.25 or 0.25. Turning it up did not produce any obvious change in the model's writing.
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.
-
00-lb load would use the following calculation:⏎[(4,000 lbs + 6,000 lbs) × 1.25] ÷ 4 = 3,125 lbs⏎According to
-
rod measuring M8-1.25 x 1 Meter. The M8 (8mm) is the rod's diameter, the 1.25 means that there is a 1.25mm distance between each thread,
-
little indenting of the skin because the device was in the shape of the finger.⏎- The Tenderlett Jr. – The incision is 1.25mm and you would know it by its white case and blue trigger. This lanc
-
poverty and insufficient production of food.⏎- one of the poorest countries in the world – in some of its regions, people live off less than 1.25 USD per day⏎- due to the food deficit, the country has to import
-
,500 miles from North Minnesota to the Gulf of Mexico. It has 250 tributaries and its flood plain covers 1.25 million square miles. Consequently, in times of heavy rainfall, this land is prone
-
793 tons of hay;⏎876,771 pounds of butter; 817,149 of cheese; 625,594 of wool; and 205,263
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 thinking about lately. I have been thinking about lately. I have been thinking about lately. I have been thinking about lately. I have been thinking about lately. I have been thinking about lately. I have been thinking about lately
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
ЛЬНАЯ РАБОТА ПО АНГЛИЙСКОМУ ЯЗЫКУ⏎КОНТРОЛЬНАЯ РАБОТА П
-
's flight. When the archer hears his shaft hiss through the air with a force not to be equalled by those of his competitors, he feels justly proud of his superior manhood. His trained subjects no one to
-
law mirroring any modern-day gun regulation, even for weapons that the people of that time could not have imagined existing?⏎Justice Breyer, joined by Justices Sotomayor and Kagan, focuses his dissent on the patent ludicrous
-
such as Enchanter (Marc Blank and Dave Lebling; 1983; ZIL), which profoundly shaped the handling of magic in IF, imported standard devices from D
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
- L1:1489 excites
- L1:32 suppresses
- L1:62 suppresses
- L0:564 weakly suppresses
- L1:773 weakly suppresses
- L0:1857 weakly excites
- L1:1042 weakly excites
- L0:1295 weakly suppresses
- L0:418 weakly suppresses
- L0:2939 weakly excites
- L1:388 weakly suppresses
- L0:1301 weakly suppresses
- L0:1150 weakly excites
- L0:2793 weakly excites
- L1:1887 weakly excites
- L0:1224 weakly suppresses
- L0:1811 weakly excites
- L0:1845 weakly suppresses
- L0:2216 weakly suppresses
- L1:331 weakly suppresses
- L1:200 weakly suppresses
- L0:345 weakly excites
- L1:19 weakly suppresses
- L0:1033 weakly excites
Feeds later layers
- L4:2938 excites
- L3:1903 weakly excites
- L7:1591 weakly suppresses
- L3:1076 weakly suppresses
- L3:1708 weakly suppresses
- L19:2084 weakly excites
- L26:527 weakly excites
- L3:793 weakly suppresses
- L21:1577 weakly excites
- L4:1982 weakly suppresses
- L10:685 weakly suppresses
- L9:1793 weakly excites
- L3:972 weakly suppresses
- L15:651 weakly suppresses
- L26:3042 weakly suppresses
- L11:2591 weakly suppresses
- L17:2105 weakly excites
- L7:1728 weakly excites
- L26:1430 weakly suppresses
- L18:2100 weakly excites
- L4:3008 weakly suppresses
- L21:2224 weakly excites
- L5:3060 weakly excites
- L11:2223 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
- ederal
- �от
- ựa
- القوم
- strengthened
- Zuk
- olon
- ymes
- ро
- Casc
Suppresses
- room
- icial
- 斯基
- ствие
- spectator
- nda
- sé
- stant
- variants
- house