'a' before few or variety
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. 24 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 'a' when it comes right before a quantity phrase like 'a few', 'a variety' or 'a host', as in 'There are a few ways'. Turning it up made the model get stuck repeating 'a number', which fits the same counting idea.
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.
-
be light brown or grayish in color and have a smooth, almost powdery texture.⏎If you find that your compost pile is not finishing, there are a few things you can do to speed up the process. One is to turn the
-
of specialists, but all will be able to diagnose and recommend treatment for TMJ.⏎How Are Temporomandibular Joint Disorders Treated?⏎There are a variety of treatment options for TMJ disorder, including physical therapy and surgery. Physical
-
less than younger individuals, losing your appetite altogether could be a sign of a more serious medical condition.⏎What Causes Loss of Appetite in Seniors?⏎There are a variety of reasons why seniors may have decreased appetites. The non-medical causes
-
sleep hygiene. Read on to learn more.⏎Negative Effects of Poor Sleep⏎Aside from feeling cranky, tired and generally out of sorts, sleep deprivation can cause a host of physical issues. Loss of sleep can affect everything from short-term memory to
-
OCT scan of the eye, your doctor can detect signs of elevation, fluid accumulation, scarring, or membrane formation.⏎Treatment for retinal disease:⏎There are a variety of<|endoftext|>In the early 90s hip-hop artist Doug E.
-
, which was previously not desirable, is now often intentionally recreated during post-processing with software like Photoshop.⏎How to Create a Vignette in Photoshop⏎There are a few ways to create a vignette in Photoshop but there are a few bits of
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 interest interest interest interest interest interest interest interest interest interest interest interest interest interest interest interest interest interest interest interest interest interest interest interest interest interest interest interest interest interest interest interest interest interest interest interest interest interest interest interest interest interest interest interest interest interest interest
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
response shaping to prevent Alzheimer's-related pathologies and memory deficits," said William Bowers, associate professor of neurology and of microbiology and immunology at the Medical Center and lead author of the article. "The vaccinated mice not only performed better
-
0 – 14 August 1958) was a French physicist, who shared the 1935 Nobel Prize in Chemistry with his wife Irène Joliot-Curie for the discovery of artificial radioactivity.⏎|
-
comprehensive review of literature on air pollution and its relationship to cardiovascular diseases. The resulting scientific statement, which appears in the current issue of Circulation: Journal of the American Heart Association, concludes that air pollution is a serious public health threat and contributes to
-
unconstitutional infringement on “liberty of contract” in a 5-4 decision.⏎Justice Oliver Wendell Holmes Jr, dissenter in the case, and Roscoe Pound (Dean of Harvard Law School), expressed views around the time that were termed
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
- L19:2727 excites
- L19:2841 suppresses
- L21:822 suppresses
- L23:96 excites
- L23:1123 suppresses
- L19:2492 excites "one of" before a group
- L23:973 excites
- L23:888 suppresses
- L23:645 excites
- L23:455 excites
- L22:784 suppresses
- L3:102 suppresses
- L18:1473 suppresses
- L17:1510 weakly excites
- L18:1895 weakly suppresses
- L23:27 weakly excites
- L23:485 weakly excites
- L20:996 weakly suppresses
- L21:1647 weakly excites
- L19:1748 weakly suppresses
- L12:1532 weakly suppresses
- L10:58 weakly suppresses
- L19:1935 weakly excites
- L22:825 weakly suppresses 'for' before a span of time
Feeds later layers
- L27:974 strongly excites
- L26:861 strongly excites
- L25:1679 excites
- L27:797 excites
- L26:2152 weakly excites
- L25:646 weakly excites measurement words before a value
- L27:2140 weakly excites
- L27:709 weakly excites
- L27:231 weakly suppresses
- L25:1586 weakly excites
- L27:663 weakly excites
- L26:2323 weakly excites
- L26:2757 weakly excites
- L25:45 weakly suppresses
- L26:2742 weakly excites
- L26:484 weakly excites
- L25:1932 weakly excites
- L25:1300 weakly suppresses
- L26:855 weakly excites
- L25:886 weakly excites
- L27:1479 weakly excites
- L25:2789 weakly suppresses
- L27:1555 weakly excites
- L26:1060 weakly 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
- number
- number
- 数
- Number
- _number
- Number
- (number
- →number
- .number
- -number
Suppresses
- xico
- idge
- ooter
- ENTE
- iston
- acle
- BEGIN
- SON
- .createClass
- otos