'of' linking a noun to a larger whole
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. 13 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 responds to the word "of" when it attaches a part or member to a bigger thing, like "atmosphere of the sun" or "crew of the ship". The effect is weaker than the neighbouring "of" neuron, and turning it up or down did not produce a clear change in the text.
Written by Claude from this neuron's activations and steering outputs. Confidence: low. 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.
-
writer at Scientific American covering paleontology, archaeology and life sciences. Follow on Twitter⏎PALEO DIET: Analyses of tartar on the teeth of Australopithecus sediba show that this early human species ate bark and
-
using images from its Large Angle and Spectrometric Coronagraph (LASCO) instrument. LASCO is used to observe the faint, multimillion-degree outer atmosphere of the sun, called the corona. A disk in the instrument is used to
-
their atmospheres. The recent discovery by the Webb Telescope, notably thanks to the NIRSpec instrument, built by Airbus, of methane and carbon dioxide in the atmosphere of K2-18b, an exoplanet 8.6 times
-
must turn sideways to squeeze between them.⏎Most of the fishermen who sail the pirogues are not from Joal. Take, for example, the crew of the 69-foot-long Mansor Sakho (she is painted sky blue
-
September, students at the New Albany club were selected to pilot the program locally.⏎Time Warner Cable — parent company of Insight Communications — partnered with Boys & Girls Club of America and FIRST Robotics to bring the program to 20 chapters across the country
-
plain why these words were put in the following sequence:⏎moral, friend, port, feet, race⏎AnswerBy adding the sound of the first five letters of the alphabet to the beginning of each word respectively, another word can be enunc
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 had 200,000,000,000,000,000,000,000,000,00,0,0,0
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
theria. If you are infected, you may infect others for up to 4 weeks. Take precautions, even if you do not have symptoms of diphtheria. Do not let anyone use your drinking cup. Everyone who lives with you
-
that is transmitted.⏎There are two different types of Low E coatings: passive Low E coatings and solar control Low E coatings.⏎The simplest way to see if your windows haveLow E glass is to hold a lit flame (match, lighter) next
-
/a0 and a0 is the Bohr radius (5.29 × 10-11 m). Calculate the position of the nodes for this wave function. Write the expected electron configurations for the following atoms: Sc, Fe,
-
01,1 requires a substantial reduction in emissions from all heavy-duty vehicle engines by 2007 through a two-pronged approach-ultra-low sulfur diesel fuel and exhaust treatment systems. While clean diesel refiners are still working on
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
- L12:1653 suppresses
- L12:2611 excites
- L6:1182 suppresses "of" after a noun phrase
- L11:1400 excites
- L11:2387 weakly suppresses
- L9:931 weakly suppresses
- L12:1301 weakly excites
- L10:574 weakly excites
- L11:1396 weakly suppresses
- L1:2789 weakly suppresses
- L12:2416 weakly excites
- L12:850 weakly excites
- L12:626 weakly excites
- L11:98 weakly suppresses
- L11:3062 weakly suppresses
- L11:1340 weakly suppresses
- L11:556 weakly excites
- L9:1909 weakly excites
- L9:884 weakly excites
- L12:3035 weakly excites
- L12:1738 weakly excites
- L11:1731 weakly excites
- L11:2155 weakly excites
- L8:429 weakly excites
Feeds later layers
- L14:1191 suppresses
- L14:2861 suppresses 'of' after a percentage
- L14:2241 suppresses
- L14:1317 suppresses
- L15:2811 suppresses
- L14:2694 suppresses
- L18:2370 suppresses
- L19:241 excites
- L19:1985 weakly suppresses
- L16:691 weakly excites
- L15:2998 weakly suppresses
- L15:1877 weakly suppresses
- L17:900 weakly excites
- L15:2877 weakly suppresses
- L14:53 weakly excites
- L15:1254 weakly suppresses
- L16:1987 weakly suppresses
- L14:630 weakly suppresses
- L18:2371 weakly suppresses
- L15:1310 weakly suppresses
- L14:2025 weakly suppresses
- L19:1742 weakly suppresses
- L16:2430 weakly excites
- L17:683 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
- далеко
- kennenlernen
- orro
- люб
- psc
- getAttribute
- mxArray
- panion
- стал
- titleLabel
Suppresses
- case
- case
- (
- CAS
- G
- 当事
- 驹
- =
- #
- غر