'a' after prepositions like 'for' or 'by'
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. 7 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 the small word "a" when it comes after prepositions like "for", "by", or "of", as in "hit by a car" or "for a while". Its negative side fires on "is" and "'s" inside longer noun phrases. Turning it up gave repetitive or unrelated text, so the steering does not confirm the 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.
-
—a 500-meter stretch of it, anyway—transforms. Its markings glow in the dark.⏎It's an experiment that's been in the works for a couple of years at this point—an attempt to solve the problem of too-dark roads
-
are a testament to the brutality of Afghan religious fundamentalists. In the case of Palmyra,<|endoftext|>Let's admit it- we all try losing weight at some point in our lives. The primary step towards shedding those extra kilos is controlling
-
(Dubbele.com; Gr 9 Up; Free) app, which will bring users to the website. The site, which has been tracking elections for a number of years, includes detailed maps and
-
English teacher, is really nice, is going to give me a letter for college.⏎- The boy, that was playing in the street, that got hit by a car was my brother.⏎- My neighbor, that lives on the 2nd
-
, 2016<|endoftext|>What is Eczema?⏎Eczema is a form of dermatitis. The term Eczema, is given to a range of persistent skin conditions. Eczema is a recurring, non-infect
-
Vein treatment⏎- Scar removal⏎- Skin tightening⏎Cosmetic dermatology - Surgical procedures⏎Most dermatologists and dermasurgeons also offer treatments for a variety of medical skin conditions, in addition to cosmetic treatments. Although most skin problems
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: do you know a person well? I have the opportunity to be involved in your life. I am very interested in what the society, the country, and the people in it think of the country." I have the opportunity to be involved in
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
2013. But Ratto and coons aren’t simply putting their designs, software and tutorials online and calling it a day: Just because something is available in theory doesn’t automatically make it accessible to all. (Linux, anyone
-
which leads to excessive consumption of critical nutrients such as added sugar, said INSP.⏎Interpreting the information presented in the GDA format to understand a product’s healthiness also requires consumers to do confusing mathematical operations, it said.⏎Undergraduate
-
species that exist in a given area. The healthier an area is, the more biodiversity it has. Different forms of wildlife and plants inhabit areas
-
and Facts about Value-Added Analysis November 11, 2011 > ˅⏎Value-added analysis uses standardized test scores to determine how much a teacher helped her students learn over the course of the year. Here are the facts
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
- L4:49 excites
- L6:2818 excites
- L5:2842 suppresses
- L2:434 excites
- L1:427 excites
- L4:19 suppresses
- L5:94 weakly suppresses
- L2:426 weakly suppresses
- L4:137 weakly excites
- L4:2386 weakly excites
- L6:54 weakly suppresses
- L6:426 weakly excites
- L6:1579 weakly suppresses
- L4:567 weakly excites
- L5:1509 weakly excites 'is' in definitions
- L3:691 weakly excites
- L3:187 weakly excites
- L3:464 weakly excites
- L2:286 weakly suppresses
- L6:1449 weakly excites
- L6:263 weakly excites
- L0:146 weakly excites
- L6:2804 weakly excites
- L3:520 weakly suppresses
Feeds later layers
- L8:868 suppresses
- L10:1897 suppresses
- L12:462 suppresses
- L8:1056 suppresses
- L9:160 excites
- L11:2108 excites
- L10:802 suppresses
- L12:2312 suppresses
- L8:323 weakly excites
- L13:2508 weakly suppresses
- L8:2193 weakly suppresses
- L21:509 weakly excites
- L9:1905 weakly suppresses
- L17:2666 weakly suppresses
- L13:1694 weakly suppresses
- L9:931 weakly suppresses
- L8:89 weakly suppresses
- L12:1458 weakly suppresses
- L10:1425 weakly suppresses
- L10:1851 weakly suppresses
- L14:400 weakly excites
- L9:1272 weakly suppresses
- L10:694 weakly suppresses
- L12:1093 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
- 之心
- ございます
- Dys
- ritos
- ernote
- STDOUT
- -Token
- -cols
- ocate
- +offset
Suppresses
- poi
- 成龙
- forms
- uses
- uses
- ondon
- unda
- passe
- must
- use