metric or imperial conversions in brackets
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. 23 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 comes right after a measurement, such as "700 kilometers (435 miles)", where a conversion to the other unit system follows. Turning it up gave text about numbers and quantities, but it does not show a clean, unique theme.
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.
-
, 2011. This image was taken through the camera’s clear filter. The distance to the surface of Vesta is 700 kilometers (435 miles) and
-
and Jeff had the privilege of naming it, and the short version is he decided it was my time.⏎The asteroid is about 1.3 km (0.8 miles) across, making it rather
-
)|⏎|• Water||83 sq mi (210 km2) 13%%|⏎|• Density||401/sq mi (155/km2)|⏎|Time zone||UTC−7 (
-
moved off their launch pads to take shelter from hurricanes and tropical storms.⏎But the shuttle cannot roll back to the VAB in winds exceeding 74 kph (46 mph), so NASA would need to move it well ahead of the storm
-
Highest point||top of flag staff|⏎|- elevation||15.75 ft (4.8 m) over stone base|⏎July 3, 19
-
50 feet (16 meters) away from the spacecraft, and had just been separated about three seconds earlier.⏎The heat shield is 15 feet (4.5 m) wide
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 an an example of a real-world work of work. Which of the following items would be a good example of work of work? (Select all of work of work items, as necessary) A. a piece of real-world work
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Indian and Atlantic oceans. The city’s elevation ranges from 5,700 to 5,930 feet (1,740 to 1,810 metres).⏎Aside from a few small streams and artificial lakes
-
, and flows generally southwestward, watering the Paris Basin, to join the Seine River at Conflans, after a course of 188 miles (302 km). It traverses the Thiérache Hills, where it
-
of beautiful new forms not seen before in inhabited structures.⏎Our final image is of the Chesa Futura in St. Moritz, Switzerland, a futuristic ski chalet which was built using many of the same automated wood-shaping technologies as are
-
⏎A typical bedroom in the rest house⏎Inside the toilet block⏎The total cost of the building construction was VND 388 Million (approximately AUD 320<|endoftext|>Go Math Worksheets – Are you aware how to optimize using numbers
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
- L22:948 strongly suppresses
- L22:1774 suppresses
- L22:2054 suppresses
- L22:2430 excites
- L22:474 excites
- L22:3034 suppresses
- L21:2040 suppresses
- L8:1245 weakly excites
- L5:967 weakly suppresses
- L19:3029 weakly excites
- L5:2760 weakly excites
- L11:2505 weakly suppresses
- L19:2275 weakly suppresses
- L21:1601 weakly suppresses
- L22:462 weakly excites
- L7:838 weakly excites
- L3:978 weakly excites
- L19:1259 weakly excites
- L21:138 weakly suppresses
- L11:2260 weakly suppresses
- L0:1932 weakly excites
- L17:2227 weakly suppresses
- L16:1183 weakly excites
- L6:1382 weakly suppresses
Feeds later layers
- L24:76 suppresses
- L24:2765 suppresses
- L26:2679 weakly suppresses
- L24:2296 weakly excites
- L25:230 weakly excites
- L24:952 weakly excites
- L24:342 weakly excites unit conversions in parentheses
- L24:2194 weakly excites
- L24:2812 weakly suppresses
- L25:2029 weakly excites
- L24:408 weakly excites
- L26:739 weakly excites
- L27:391 weakly excites
- L24:2657 weakly suppresses
- L24:2527 weakly suppresses
- L25:2519 weakly excites
- L24:2854 weakly suppresses
- L25:1749 weakly suppresses
- L26:463 weakly suppresses digits inside years and numbers
- L24:2895 weakly excites
- L27:1044 weakly suppresses
- L24:2911 weakly excites the 'cul' fragment in culprit and culminate
- L26:1698 weakly suppresses
- L24:2481 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
- mtree
- boarding
- napshot
- Double
- idden
- Double
- .hit
- recht
- .Hit
- statuses
Suppresses
- VED
- boundary
- iota
- ài
- swell
- illus
- pear
- urus
- 舌头
- Head