quantity words like 'two' and 'several'
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. 21 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 fires on number-like words such as 'two' or 'several' that come right after a verb or preposition and set up a count of things, as in 'separate two organizations'. Turning it down made the model repeat small quantity phrases like 'a few' and 'a bit', so the neuron seems tied to how much or how many.
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.
-
a drug.⏎Over the years, occasional compounds of this sort would show up in the literature, usually as part of an effort to, for example, bind to two receptors at once when they were thought to be near to
-
$1,500.⏎The same idea holds true for businesses. The difference between gross income and net income can be enormous, and the numbers can tell two wildly different stories as to the health of the business.⏎Too many small businesses focus
-
other defense proposals, threatens to further separate two organizations already experiencing strains on their relationship.⏎The EU’s shifting perspective toward defense comes
-
known as a chromosomal microarray analysis, which can detect large scale deletions and duplications of the genome.⏎Learn more about how genetic testing positively impacted two SPARK participants: Hope and Alex.⏎While these techniques were state-of-the-art
-
there are an abundance of activities children<|endoftext|>January 2022 Newsletter⏎Mosquitofish are small guppy-like fish that can eat several hundred mosquito larvae each day. The District uses mosquitofish as part of
-
that parents are concerned about.⏎4. In some urban school districts, there are very low levels of parent engagement, due to barriers that can include parents who work two jobs, kids who have foster parents or are being raised by grandparents, single parent
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: A two- a four a a a a a a a a a a a a a a a a a a a a a a a a a a a a a a a a a a a a a a a a a a a
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
with you. If we have whetted your appetite for interesting food facts, here are some previous posts you may find interesting:⏎- Furikake: Everything you need
-
berries seem to affect the genes that are related to fat-burning and storage.⏎So what does this mean for you? Toss a handful on your yogurt, throw a few in a salad or add some to your morning cereal. Even though this
-
, but we’re still trying to imitate the real world.⏎Animation on the Web: a short history⏎Before we get onto more practical things, let’s take a look at how animation on the Internet came to its present (and very cool
-
mination érrimo-aSuperlative adjectives are formed with the suffix érrimo-a, a Latin termination occurring while it is completely correct obsolete. Here are some examples: pobre-pauperrimo, celebre-celeb
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
- L19:1501 suppresses
- L19:1935 excites
- L18:2804 weakly suppresses
- L16:1333 weakly excites 'two' after 'divided into' or 'there are'
- L18:1265 weakly suppresses
- L20:36 weakly suppresses
- L19:1379 weakly excites
- L18:801 weakly excites
- L18:2021 weakly excites
- L18:2788 weakly excites
- L18:1942 weakly excites
- L18:857 weakly suppresses
- L20:2827 weakly suppresses
- L12:2053 weakly excites
- L12:2424 weakly excites
- L18:692 weakly excites
- L19:687 weakly suppresses
- L17:131 weakly suppresses
- L18:1076 weakly excites
- L7:1706 weakly suppresses
- L19:1748 weakly suppresses
- L16:1860 weakly suppresses
- L6:2583 weakly suppresses
- L10:1990 weakly excites
Feeds later layers
- L24:823 suppresses
- L27:797 suppresses
- L25:45 excites
- L25:1586 suppresses
- L24:2538 suppresses
- L27:663 suppresses
- L27:1124 suppresses
- L22:103 excites
- L22:926 suppresses
- L25:433 excites
- L26:213 excites
- L22:171 suppresses
- L25:2547 suppresses
- L23:2883 weakly suppresses
- L27:709 weakly suppresses
- L22:360 weakly suppresses
- L24:488 weakly excites
- L27:173 weakly excites
- L22:825 weakly suppresses 'for' before a span of time
- L27:231 weakly excites
- L26:484 weakly suppresses
- L26:411 weakly excites
- L25:2527 weakly suppresses
- L24:826 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
- lettes
- tings
- �재
- UDGE
- pequeño
- ovich
- arently
- lore
- bordel
- rowsers
Suppresses
- some
- a
- 一些
- Some
- some
- Some
- 几个
- 一个
- 几张
- an