'of' before a measured value
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. 20 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 word 'of' when it comes right before a measurement or quantity, like 'a depth of 12 m' or 'efficiency of 4-6%'. It switches off on words like 'a' or 'low' that lead into a number. Turning it up made the model start writing lists of numbers and percentages.
Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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.
-
options for artificial teeth replacements, the natural teeth will always be the preferred teeth of professionals.⏎By<|endoftext|>Demonstration: This rough method can produce a difference of mass of at least 8 g . With enough apparatus
-
by the authors.<|endoftext|>Image: Common Seadragon, Phyllopteryx taeniolatus⏎A male Common Seadragon at a depth of 12 m at 'two poles', off Bateau Bay, New South
-
— A thin-film, silicon photovoltaic cell having no crystalline structure. Manufactured by depositing layers of doped silicon on a substrate. Efficiency of panels is generally 4-6% and performance does degrade rapidly - as much
-
to have an asymmetry in the normal population, having a leftward bias in right-handed individuals, people who possess absolute pitch have an increased leftward asymmetry of the planum temporale. This
-
-aircraft coverage, and even sniping ability. Its combination of exceptional mobility – being light enough to be carried by one man – and high rate of fire (of up to 900 rounds per minute) was unmatched. It entered service
-
2.2% and their mean agreed with that obtained by the Kjeldahl method. By use of the prescribed dilution of the sample, detection limits of 0.1% can be achieved for all three cations. ??
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: the future of a health system for the future. It is a health system that will have no single, clear, public health sector, and that will have a more efficient and diverse, and more transparent, and more sustainable, and more efficient
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
0 to $102,202 per year with median annual wages estimated at $49,490. Hourly rates range from a low of $13.50 to a high of $49.0
-
of injections performed in the United States, estimated from Medicare procedure codes, increased from fewer than 3000 per year in 1999 to an estimated 6.5 million in 2016. This number continues
-
50 years; further research shows turtle dove numbers have plummeted by 80 per cent since 1995, hedgehog numbers are down by a third since 20
-
700 among adults ages 35 and 44, more than 5,600 among adults 45 to 54, more than 3,800 among adults 55 to 64
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
- L19:248 strongly excites
- L19:1211 strongly excites
- L19:1222 strongly suppresses
- L19:3047 suppresses
- L18:2183 suppresses
- L10:868 suppresses
- L19:1921 excites
- L17:1372 suppresses
- L19:1454 excites
- L17:2230 suppresses
- L18:390 excites
- L17:2083 excites
- L14:641 excites
- L6:1149 suppresses
- L18:2150 excites
- L7:1524 suppresses
- L16:3042 suppresses
- L14:471 excites
- L18:2384 suppresses
- L19:803 excites
- L6:1008 excites 'of' after a measured quantity
- L14:1793 suppresses
- L6:756 suppresses
- L18:291 suppresses
Feeds later layers
- L21:676 strongly excites
- L21:175 strongly excites
- L24:722 strongly excites 'about' or 'approximately' before a number
- L26:667 strongly excites
- L25:2902 strongly excites
- L25:253 strongly excites
- L24:587 excites
- L25:46 suppresses
- L23:198 excites
- L27:287 excites
- L24:126 excites
- L26:476 excites
- L21:336 excites
- L27:2172 suppresses
- L26:423 suppresses
- L22:1146 excites blank before a time or count
- L26:172 excites
- L24:162 suppresses
- L21:543 excites
- L25:2908 excites
- L27:113 suppresses
- L25:1212 excites
- L25:1136 excites
- L27:163 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
- fifty
- 五百
- 五千
- thirty
- 六十
- 三十
- forty
- 三千
- 约
- 四十
Suppresses
- eve
- enza
- ']}}</
- endphp
- güneş
- AVE
- IMP
- council
- eenth
- ilename