'but' joining contrasting clauses
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. 11 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 word 'but' when it links two contrasting ideas, such as 'a small but obvious distinction' or 'the percentage differences were minor, but still noteworthy'. Turning it up did not produce a clear theme in the continuations.
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.
-
new st<|endoftext|>Can a small camera possibly have a large lens in the front?⏎There are really "large" lenses. This search is not the biggest one, but the look of it is great.⏎But you are referring to a large diameter compared
-
have an X-Ray? Luckily, it has been over a few years since I last had one, but I always remember the strange experience seeing those X-Ray photos hung up on display.
-
of references, but its sources remain unclear because it has insufficient inline citations. (May 2009)|⏎Lexicography is divided into two separate but equally important groups:⏎- Practical lexicography is the
-
. More important, but sometimes less emphasized (pun intended), is Rhythm.” For further<|endoftext|>By Robert
-
may be a slight but obvious distinction between New Years in North Korea and with the rest of the world.
-
had increased from 8% to 84%, and that the number of reported head injuries and potentially serious head injuries decreased.⏎The percentage differences were minor, but still noteworthy.⏎A similar analysis was conducted by the NHS in the United Kingdom.
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 future of the future of the future, the future of the future, the future of the future, the future of the future, the future of the future, the future of the future, the future of the future, the future of
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
humans, activity levels vary widely from couch-potato-style inactivity to highly active athletic endeavors.⏎"We have a huge epidemic of obesity in Western society, and yet we have little understanding of what determines variation among individuals for voluntary exercise levels," Garland
-
medicine. Especially in India, Neem has been used for thousands of years.⏎Not many CPG companies in the US food industry are currently using Neem.⏎However, many hygiene products – such as
-
house quickly if a fire starts.⏎However, fewer people know how to prevent fires in a motor vehicle. Vehicle fires are among the most terrifying types of accidents, yet few people can name ways to prevent them or address them if they happen. Experienced
-
area and increase property values.⏎The students found 74 percent of people in the Millers Bay neighborhood had no knowledge of the benefits of shoreline restoration. However, after students educated them on the topic, more than 80 percent of homeowners
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
- L9:1755 suppresses
- L10:339 suppresses
- L4:659 excites
- L3:919 excites
- L6:1054 excites
- L7:1380 weakly excites
- L9:2802 weakly suppresses
- L5:2775 weakly suppresses
- L3:2553 weakly excites
- L9:2516 weakly excites
- L8:1784 weakly excites
- L0:2223 weakly excites
- L9:1579 weakly suppresses
- L10:743 weakly suppresses 'therefore' or 'thus' after a subject
- L10:3039 weakly excites
- L8:2563 weakly suppresses
- L10:2742 weakly suppresses
- L4:1319 weakly excites the 'B' in vitamin B
- L1:2957 weakly suppresses
- L5:1859 weakly excites
- L5:314 weakly excites
- L2:219 weakly excites
- L8:487 weakly excites
- L2:2706 weakly excites
Feeds later layers
- L17:1730 strongly excites
- L12:387 excites
- L15:1791 excites
- L12:2005 excites
- L16:1488 excites
- L15:1422 excites
- L12:2622 excites
- L15:1948 excites
- L18:1682 weakly suppresses
- L19:2213 weakly excites
- L16:2408 weakly excites
- L14:2621 weakly excites
- L17:2269 weakly suppresses
- L24:1976 weakly excites
- L17:2707 weakly excites
- L12:2965 weakly suppresses
- L15:1345 weakly excites
- L17:2081 weakly excites
- L16:1470 weakly excites
- L17:337 weakly suppresses
- L16:2212 weakly excites
- L25:2067 weakly suppresses
- L20:1684 weakly excites
- L20:1908 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
- ilda
- efs
- Mathematical
- lef
- hrs
- emd
- ämpfe
- íveis
- 并不代表
- lige
Suppresses
- asti
- ylon
- gtk
- ?”
- atism
- izione
- .DriverManager
- fl
- !”
- 或者