'by' giving the amount of a difference
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. 14 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 "by" when it introduces how much something changed or differed, as in "predates the Altair by almost two years" or "shifted in the same direction by the same amount." When turned up, the model's writing tended toward numbers and amounts, which fits this pattern.
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.
-
well as at the fueling station, which also is being built and operated in partnership with NW Natural.⏎Trucks running on natural gas cut smog-producing pollutants by up
-
wrecks. At nearly 3,000 years old, this Bronze Age find is particularly significant, not just because it predates the birth of Christ by a millennium, but because in addition to the weapons and the precious gold jewelry,
-
recording of a human voice.⏎The recording played Thursday predates Thomas Edison's invention of the phonograph (previously thought to have recorded the first sound) by 17 years. It captured about 10 seconds of the French folks
-
eter. If the voltmeter is not correctly zeroed before use, then all values measured by the voltmeter will be biased, i.e., offset by some constant amount or factor. However, even if the utmost care is taken in
-
employed the Intel 4004 CPU, were the first American microcomputers ever to be sold and pre-date the Altair 8800 by almost two years.⏎The Intellec Microcomputer Development Systems (MDS)
-
characters entered by the user, but in reverse order. The second string contains the 5 characters entered by the user, but whose ASCII code has been decremented by
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 question that I think I can be a person who is as strong as the strongest person who is the strongest person who is... the strongest one who... the strongest one who... the strongest one... the strongest. ... the strongest... the
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
patient.⏎Medical-Surgical Nursing at a Glance is ideal for nursing students looking for a clear and visual summary of care of the adult patient.⏎Results 1-5 of 41⏎The nurse practitioner is skilled at
-
(6-8) Show all Grade Level⏎Resource Type: Activities⏎Results 1 - 7 of 7 matches⏎From Grid to Home⏎Lisa Gardiner, National
-
Necessary intensities are in the range between 1010- 1011W/cm2.1, 2 at such intensities, gas molecules
-
V | W | X | Y | Z | 0-9⏎21 results for Schools: |⏎See previous results⏎See results 11 - 15⏎View all
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
- L12:2545 suppresses
- L6:756 suppresses
- L10:868 suppresses
- L10:1214 suppresses
- L13:884 excites
- L5:685 suppresses
- L13:1308 excites
- L6:1452 weakly suppresses
- L3:2093 weakly excites
- L7:1088 weakly suppresses
- L12:1777 weakly suppresses
- L13:513 weakly excites
- L13:2770 weakly suppresses
- L12:2801 weakly excites percentages and time spans in statistics
- L0:2932 weakly suppresses
- L9:1108 weakly suppresses
- L11:2493 weakly suppresses
- L13:2915 weakly suppresses
- L9:2078 weakly suppresses
- L2:1045 weakly excites
- L4:1672 weakly suppresses
- L12:740 weakly excites
- L11:95 weakly excites
- L8:2164 weakly excites
Feeds later layers
- L18:3014 strongly excites
- L16:3022 excites
- L17:2721 excites
- L16:2032 excites
- L15:2091 excites
- L15:3028 excites
- L24:722 excites 'about' or 'approximately' before a number
- L24:587 excites
- L15:1909 excites blank slot before a number
- L19:1419 excites
- L17:222 excites
- L17:1951 excites
- L21:175 excites
- L17:1570 excites
- L17:1804 excites
- L17:1504 excites
- L18:291 suppresses
- L17:2952 excites
- L17:2208 excites
- L17:1512 excites
- L16:1775 excites
- L18:2976 excites
- L19:1458 excites
- L19:2151 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
- 一刀
- 一秒
- 几分钟
- 一个小时
- 多久
- 半个
- few
- 一点点
- 一步
- -half
Suppresses
- sville
- urai
- jec
- juxtap
- uß
- surrounded
- antasy
- 周三
- ypad
- sounded