unclear: word before a twist or contrast
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. 17 of Qwen3-0.6B. Its strongest snippets are mixed: they peak on different tokens. 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 a word in the middle of a sentence where the writing seems to be setting up a turn, like 'far from over', 'doesn't end there' or 'wasn't the first'. It is hard to pin down more precisely. Turning it up made the model's continuations wander off into unrelated topics.
Written by Claude from this neuron's activations and steering outputs. Confidence: low. 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.
-
technology status, the fuel cell has experienced an astounding development. Fuel cells are already being used in a variety of applications today. But its impressive career is far from over. Because of their simple operation, the use of fuel cells in electric vehicles represents
-
Dell Encrypt?⏎Dell encrypts all personal information you give us when you place an order online. Remember, if you feel uncomfortable providing any of this information online, please feel free to call one of our representatives at 1-800
-
a 10-fold increase in storage capacity as well as a 300x improvement in performance over today’s drives.⏎The good news doesn’t end there, though. Even though this new technique relies on heat, it is thought to
-
year of learning, students often spend their summers not thinking about schoolwork and lose a good amount of the learning progress they have made. At Schoolmart, we understand the importance of continuous learning and are determined to bridge the summer knowledge gap with helpful
-
famous Paul who is credited with writing much of the New Testament.)⏎Although all the original 12 apostles are important, some get top billing. Here’s why:⏎
-
were less hardy than their parents. The same can be said about Dolly. She aged prematurely and passed away after 6 years.⏎Dolly wasn’t the first vertebrate to be cloned. That honour belongs to a carp, cloned in
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 first time, you have to be careful with the first time you have to be careful with the first time you have to be careful with the first time you have to be careful with the first time you have to be careful with the first
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
have to try slow your heart down so you can hit a target that is about the size of a coin. I don't know how they do it. It sounds really difficult.⏎Racers wear cross-country skis, ski
-
to read. Reading is fascinating. It's full of surprises. And it will take you places you can't get to in any other way.⏎- Do EACH of the following:⏎- Learn how
-
can be as simple as a long alkyl chain, or a short stretch of ethylene glycol ethers, up to as complex a linker as you have the patience to search for.⏎It seems like a rather simple idea that you could get two
-
of Frank Churchill’s secret engagement to Jane.⏎“Good God!” cried Emma, “Jane actually on the point of going as governess! What could he mean by such horrible indelicacy? To suffer her to engage herself
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
- L15:484 excites
- L15:348 suppresses
- L10:2688 suppresses
- L10:1610 excites
- L13:2710 suppresses
- L10:976 suppresses
- L14:1580 excites
- L10:375 excites
- L15:127 excites
- L16:1308 excites
- L10:380 excites
- L12:2444 excites
- L15:1803 excites
- L5:374 excites
- L10:1320 suppresses
- L9:1234 excites
- L11:772 excites
- L12:1831 weakly excites
- L14:3041 weakly excites
- L10:1289 weakly excites
- L14:434 weakly excites
- L11:961 weakly excites
- L16:632 weakly suppresses
- L8:2370 weakly suppresses
Feeds later layers
- L19:1653 excites
- L21:1972 weakly excites
- L18:2050 weakly excites
- L18:2722 weakly excites
- L18:2636 weakly excites
- L18:2704 weakly excites
- L22:2880 weakly excites
- L23:1227 weakly excites
- L19:1620 weakly excites
- L25:2960 weakly excites
- L22:705 weakly excites
- L20:2470 weakly suppresses
- L21:2358 weakly excites
- L19:585 weakly suppresses
- L19:3047 weakly excites
- L19:1868 weakly excites
- L21:1945 weakly excites
- L18:2545 weakly excites
- L20:2265 weakly suppresses
- L21:679 weakly excites
- L22:1390 weakly excites
- L19:2552 weakly excites
- L20:802 weakly excites
- L23:2716 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
- ois
- '`
- iores
- addock
- UDIO
- aised
- ValidateAntiForgeryToken
- reetings
- てくれ
- '!
Suppresses
- 下午
- бан
- *ft
- erg
- .cy
- ification
- -ph
- bus
- scheduler
- 通过