"from" after change or transfer verbs
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. 5 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 lights up on the word "from" when it follows a verb of moving, switching, or changing, such as "switch from," "transform from," or "translate from." Its negative side is very weak and not clearly meaningful. Turning it up did not produce text that clearly showed 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.
-
control circuit continuously monitors the battery's voltage. When the battery is fully charged, the circuit automatically turns on a power inverter and switches the appliance from running on grid power to running on the energy stored in the battery. Then when
-
led in a quiet residential neighborhood in Long Beach, Calif., are 4.7 acres of Spanish, Mexican and American history, where families helped transform southern California from its ranching beginnings to a modern, urban society.⏎Rancho Los C
-
serious consideration. Now, take a look further for hacks that will turn you from non-native to native English writing prowess.⏎- Like from a speaking perspective This doesn
-
28, providing they can switch their children from a 25th percentile teacher to a 75th percentile teacher.⏎
-
few seconds:⏎See how “normal” is translated from English to French with more examples in context⏎- Create your own vocabulary list⏎
-
of snack manufacturer Snyder’s-Lance Inc. illustrates how this can be applied to a major change initiative. The company’s CEO initiated a major organizational change to move from a more traditional, top-down culture to a servant-leadership culture. A
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: you are being asked to answer a question, and you are trying to figure out what it is. a, i, s, c, o, l, f, m, c, r, d, c, h, r,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
temperature)⏎and E is the flux in W/m²⏎More about this formula and background on the material properties in Planck, Stefan
-
INTON LOCKHART, Ph. D., LL. D., Author of Laws of
-
/or data.⏎Some local statistics:⏎- Smart phone penetration around 51%⏎- >80% use cellphones to communicate⏎- 4G/LTE subscriptions 22%⏎- 4G/LTE coverage 86%⏎
-
abyte data transfers every second, initially it is expected the technique will allow data to be stored at a density of 10 petabytes/m2 and using a data transfer rate of 200Gb/s. That’s
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
- L3:1858 excites
- L3:1907 weakly suppresses
- L1:765 weakly excites
- L4:1874 weakly suppresses
- L1:2454 weakly excites
- L4:600 weakly suppresses
- L2:1501 weakly suppresses
- L3:2458 weakly excites
- L2:237 weakly excites
- L3:1001 weakly excites
- L1:2863 weakly excites
- L1:1608 weakly excites
- L1:1200 weakly suppresses
- L4:2111 weakly suppresses
- L3:2149 weakly suppresses "that" after demand or recommend verbs
- L4:1635 weakly excites 'light' as an adjective (pale, not heavy)
- L4:2649 weakly excites
- L0:1270 weakly excites
- L1:1480 weakly suppresses
- L2:2427 weakly excites
- L3:1787 weakly excites
- L3:1982 weakly excites
- L0:920 weakly suppresses digit '9' inside years and numbers
- L1:2393 weakly suppresses
Feeds later layers
- L8:2018 excites
- L9:229 excites
- L18:2272 excites
- L14:513 weakly excites
- L14:2475 weakly excites
- L17:1488 weakly excites
- L12:1037 weakly suppresses
- L18:1857 weakly suppresses
- L13:2762 weakly excites
- L17:2683 weakly suppresses
- L16:1779 weakly suppresses
- L26:1675 weakly suppresses
- L16:1514 weakly suppresses
- L22:2582 weakly excites
- L24:1070 weakly excites
- L6:1168 weakly excites
- L6:364 weakly excites
- L7:2179 weakly suppresses
- L23:2868 weakly suppresses
- L25:454 weakly suppresses
- L9:723 weakly suppresses
- L23:1660 weakly excites
- L24:2083 weakly suppresses
- L27:1717 weakly suppresses 'st' starting a word fragment
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
- 换了
- /question
- iated
- aque
- rawn
- ftware
- ardi
- 腑
- hdr
- 荷
Suppresses
- )[-
- pop
- отов
- IMPLIED
- inness
- pop
- еть
- etimes
- CrossRef
- houses