unit conversions in parentheses
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. 24 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 opening bracket that follows a measurement, such as '6,962 m (22,841 ft)' or '9 inches (23 centimeters)'. It is the bracket that starts a conversion into other units. Turning it up made the writing break into number-heavy and unclear text, so the steering shows nothing about this pattern.
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.
-
are the highest mountains of each of the seven continents.⏎Aconcagua, South America⏎With its cone-shaped peak rising at 6,962 m (22,841 ft), Cerro Aconcagua holds the title
-
the image has a resolution of about 70 meters (230 feet) per pixel. This image was acquired during the HAM
-
9.6 km/h)|⏎|Range:||3,500 nautical miles (6,482 km) at 12 knots (22.2
-
Visayan warty pigs seasonally grow large manes that start at the top of their head and travel partially down their back and grow up to 9 inches (23 centimeters) long! The mama of the group is Kurit (
-
|Dimensions||H: 11 1/16 in. (28.2 cm.), W: 8 7/8 in. (22.5 cm.)|⏎|Location||Alte Pinakothek
-
990|⏎|-||Restored||August 21, 1991|⏎|-||Total||64,589 km2 (124th)⏎24,938 sq mi⏎|-
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: "11111111111111111111111111111111111111111111111
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
40), the first of three volumes on the selected writings of respected English Romantics Anna Laetitia Barbauld and her brother, John Aikin, Sarah Hale explains that the
-
of carbon should be kept low for industrial and machinery lubrication.⏎* Malaysian Palm Oil Board (MPOB),⏎6 Persiaran Institusi, Bandar Baru Bangi,⏎43000 Kajang, Selangor
-
Generally the ticks that cause Lyme Disease in the USA carry the Borrelia Burgdorferi bacteria in their gut, whilst in Europe it is often the Borrelia Afzelii; these bacteria cause Lyme disease. Currently research
-
attempt by society members in 1897 to impose on her an alternative dress code to the customary wig and gown, her exclusion from the society's annual Bar dinners, and the<|endoftext|>In Year 4, our class text this half term
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
- L22:1795 suppresses
- L22:2327 excites
- L21:879 suppresses
- L22:1573 excites
- L23:2720 excites
- L22:1190 excites
- L22:854 excites
- L19:2609 suppresses
- L22:1432 excites
- L21:1378 suppresses
- L20:256 excites
- L18:385 weakly excites
- L23:554 weakly suppresses
- L23:2728 weakly excites second digit of years like 2020
- L10:2719 weakly suppresses
- L23:808 weakly suppresses
- L21:2586 weakly excites
- L23:2035 weakly excites
- L17:2227 weakly suppresses
- L23:2760 weakly suppresses
- L22:176 weakly suppresses
- L23:1272 weakly excites
- L23:723 weakly excites metric or imperial conversions in brackets
- L23:399 weakly suppresses
Feeds later layers
- L27:189 suppresses
- L26:270 suppresses
- L26:834 excites
- L27:352 excites
- L25:1749 suppresses
- L25:379 excites
- L26:463 suppresses digits inside years and numbers
- L27:678 weakly suppresses
- L25:2298 weakly excites
- L27:364 weakly suppresses
- L25:2519 weakly suppresses
- L25:1452 weakly suppresses
- L26:598 weakly suppresses spot just before a year or century
- L25:492 weakly suppresses
- L25:145 weakly suppresses
- L27:992 weakly excites
- L27:2706 weakly suppresses
- L25:1287 weakly suppresses
- L26:1734 weakly suppresses 'the/as' before 'following'
- L25:621 weakly suppresses
- L26:2648 weakly suppresses
- L25:1322 weakly suppresses
- L25:1606 weakly suppresses
- L26:556 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
- yscale
- lığın
- علامات
- marked
- ики
- eworthy
- :+:
- aller
- Microwave
- 维奇
Suppresses
- bell
- LEE
- ront
- .Sp
- incorpor
- PERT
- .SP
- 涌现
- oreal
- son