inside numbers and citations
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. 16 of Qwen3-0.6B. Its strongest snippets are mixed: they peak on different tokens. 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 pieces of numbers, dates, and reference details, such as a digit in a postcode, a dot in an author initial, or the dollar sign before a sum. It goes negative on dashes in year ranges and commas in plant names. Turning it up made the model drift into number lists.
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.
-
references: Gwartney, J. D., Stroup, R. L., Sobel, R. S., & Macpherson, D. A. (2018). Economics: Private⏎and public choice (16
-
Water-Quality Assessment (NAWQA) Program of the U.S. Geological Survey (USGS).1 Chloroform and other trihalomethanes, the most commonly detected compounds, were found in about 9 percent of the sampled
-
570 and at Hammocks Beach State Park located at 1572 State Rd 1511, Swansboro, NC 28584.⏎The Coastal Federation’s Recycling for Reefs Program collects shells and
-
in reality is brown, colors were produced by putting together thousands of of NASA’s images of Mercury surface, Space Environment, Geochemistry, and Ranging (Known as Messenger) spacecraft, which is in orbit around the innermost planet. Engineers used
-
2025 and 2030. Respectively, these targets are 15% and 37.5% lower than the 2021 figure. The targets were finalized in 2019
-
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 collects
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 number of a few minutes, the number of a few minutes, the number of a few minutes, the number of a few minutes, the number of a few minutes, and, in some way, the number of a few minutes,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
, including handling of certain kinds of global variables.⏎The first step in creating a GLib shared library is to create a ``static'' library (also called an ar archive).
-
results⏎Authors: Walker, Robert J. (John), 1801-1869⏎Date: May 27, 185
-
for interior growing conditions. It also likes to be potbound and will bloom even in a relatively tiny pot.⏎Common name: Fire lily, bush lily, clivia⏎Botanical name: Clivia miniata⏎Flowers: Rounded
-
and examples of With clause in oracle let us first check some important bullet points of With Clause :⏎- With Clause in Oracle is released in Oracle 9i release 2 to improve the performance of complex sql queries.⏎- The clause works like a
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
- L15:7 excites
- L14:597 suppresses unclear: document ends and list breaks
- L12:1014 suppresses
- L13:2302 suppresses
- L15:73 weakly excites
- L8:1219 weakly suppresses
- L15:370 weakly suppresses
- L14:1915 weakly suppresses
- L12:730 weakly suppresses
- L13:1345 weakly suppresses
- L10:1731 weakly excites
- L5:84 weakly excites
- L5:75 weakly suppresses
- L8:177 weakly excites
- L3:533 weakly suppresses
- L2:49 weakly excites
- L2:1537 weakly suppresses
- L11:1231 weakly suppresses
- L5:71 weakly suppresses
- L8:379 weakly excites
- L10:1700 weakly excites
- L7:1265 weakly suppresses
- L12:229 weakly suppresses
- L4:1647 weakly excites
Feeds later layers
- L19:2731 weakly excites
- L23:1772 weakly suppresses
- L20:1417 weakly excites
- L21:1724 weakly suppresses
- L18:2235 weakly suppresses
- L19:752 weakly suppresses
- L20:1448 weakly excites "the" in mid-sentence, weak
- L27:1605 weakly excites
- L17:2646 weakly suppresses
- L17:1088 weakly excites
- L22:1349 weakly excites
- L20:2133 weakly suppresses letters inside spelled-out words
- L19:1925 weakly excites
- L17:1397 weakly suppresses
- L19:2225 weakly suppresses
- L20:2510 weakly suppresses
- L19:2030 weakly suppresses
- L27:235 weakly suppresses
- L24:183 weakly excites
- L23:2817 weakly excites the word "no" in "no longer"
- L18:2347 weakly excites
- L17:497 weakly suppresses
- L22:621 weakly excites
- L23:247 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
- Strauss
- idar
- NavController
- tin
- stabilization
- arity
- /*!<
- +:+
- interfaces
- strt
Suppresses
- OMET
- ount
- omet
- adox
- leg
- 簿
- oles
- 忘了
- ちゃんと
- 울