the last digit of decade years
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. 6 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 final zero in decade years like 1970s, 1980s or 1930s, so it is reacting to the point in a sentence where a decade is named. Turning it up pushed the model toward unrelated maths word problems and did not make it talk about history.
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.
-
and Quantum Communication.⏎How it started:⏎Artificial Intelligence started in the 1950s, but its importance was recognized in the 1970s.⏎Japan was the first to take this initiative and in 198
-
mass incarceration of Japanese Americans, Asian Americans called for affirmative action programs for racially excluded minorities. From the 1960s through the 1980s, Asian Americans helped force higher education open for excluded communities of color and benefited
-
insights into the Martian past demonstrated the science potential of investing in imaging of the planets from orbit. By the end of their missions in the early 1980s, the Viking landers and orbiters had accumulated over 50,
-
resulted in a period when nearly every American family had a shot at a better life.⏎Our recent history tells a different story. Since the early 1980
-
⏎It became particularly strong during World War II--when advancing Japanese forces cut off U.S. access to hevea--and during the 1970s.⏎Guayule grows much faster than hevea, but the latter
-
Norman Davies noted that American forces fought with distinction on two fronts, but that it was our home factories that turned the tide. In the late 1930's, American industry produced consumer goods at a dog's pace. By 1
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: "A lot of things." The number of the blog post is a great place for the word 'something' in the context of a blog post. The first thing to note is that the word 'something' is used in a wide range
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
etal force acting on a car of mass 2000 Kg turning around a circular bend of radius 50 meters at the constant speed of 36 km/hr?⏎A) 4,000 N⏎B)
-
) with a reasonable profit at the present price of crude oil, and 2,500 barrels per acre for an area of 2,248 acres (5,620,000 barrels total), with no
-
optical mice as well as 800 DPI laser mice are very popular, but some high-end models are capable of 1000, 1600 or even 2000 DPI tracking speeds.⏎The Personal System
-
drop for different weight bullets ranging from 250, 325, and 405 grains.⏎This following table is a ballistics
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
- L5:1641 suppresses
- L5:1072 weakly excites
- L4:12 weakly suppresses
- L4:2492 weakly excites
- L4:2888 weakly suppresses
- L5:881 weakly excites
- L4:245 weakly excites
- L2:1605 weakly suppresses
- L2:208 weakly excites
- L3:612 weakly excites
- L1:1254 weakly excites
- L5:2154 weakly excites
- L5:28 weakly excites
- L4:1996 weakly excites
- L1:462 weakly suppresses
- L3:2980 weakly suppresses
- L0:2023 weakly suppresses
- L3:1678 weakly suppresses
- L5:2513 weakly excites
- L5:2937 weakly excites
- L5:2727 weakly excites
- L2:2330 weakly suppresses
- L3:512 weakly suppresses
- L3:176 weakly suppresses
Feeds later layers
- L7:1808 strongly excites
- L24:1676 excites zeros in large round numbers
- L27:2476 weakly suppresses
- L17:409 weakly excites
- L8:1402 weakly excites
- L13:2379 weakly excites the digit zero in numbers
- L9:977 weakly excites the last digit of a decade year
- L22:687 weakly suppresses
- L20:854 weakly suppresses
- L17:2270 weakly suppresses
- L15:1965 weakly suppresses
- L14:2970 weakly excites
- L25:2616 weakly suppresses
- L19:623 weakly suppresses
- L7:1696 weakly excites
- L21:474 weakly suppresses
- L22:2300 weakly suppresses
- L27:2740 weakly suppresses
- L17:1466 weakly excites
- L26:847 weakly suppresses
- L13:2708 weakly suppresses
- L17:2305 weakly suppresses
- L10:1792 weakly suppresses
- L23:2406 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
- awning
- awns
- ��
- quartered
- ��
- licted
- elps
- assador
- チェ
- ��
Suppresses
- да
- атель
- ale
- опы
- Lock
- immer
- дал
- 使用
- Use
- 输卵管