advice words like avoid, limit, recommend
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. 13 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 advice words such as 'avoid', 'limit' and 'recommend', especially in phrases like 'Foods to avoid' or 'It's best to avoid'. Turning it up did not bring out that advice theme; the model wrote repetitive or unrelated text.
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.
-
.⏎But like I said before: a ferret cannot digest plant matter which means that giving them anything other than meat will lead to health problems.⏎Foods to avoid:⏎-
-
out for sure ingredients. Here's an inventory of toxic foods you would like to avoid.⏎- Chocolate is one among a lot of renowned toxic human food a dog will
-
ings.⏎4. Limit sugary and acidic foods or drinks:⏎Sugary and acidic foods and beverages can erode your tooth enamel and cause cavities. Limit your intake of sugary or acidic foods and beverages, and drink plenty of water
-
would like to support the mission of Open Culture, consider making a donation to our site. It’s hard to rely 100% on ads, and your contributions will help us continue providing the best free cultural and educational materials to learners everywhere
-
the fitness industry. As does the “no pain, no gain” myth, which came to prominence in the early 1980
-
tires and who overdrive the conditions.<|endoftext|>Viewing page 28 of 101⏎It looks like you're using a mobile device. We recommend using a physical keyboard for transcription entry.⏎28 REPORT NATIONAL ADVISORY COMMIT
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: I have been thinking about the things I have been thinking about the things I have been thinking about the things I have been thinking about. I have been thinking about the things I have been thinking about the things I have been thinking about the things
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
presents both an opportunity and a challenge for the planet. Carlos Barria / Reuters⏎A nurse takes care of newborn
-
made their way through the “land,” they count their tokens and see if they made it through correctly by solving another addition problem at the end. This is a great self-checking component of the activity.⏎SUBTRACTION FACT LAND HELPS STUD
-
u-substitution twice (second time with w)⏎⇐ Use this menu to view and help create subtitles for this video in many different languages. You'll probably want to hide YouTube's captions if using these subtitles.⏎- Let's see if we can
-
part of your educational outreach around Earth Day, or use it as part of your lesson—it’s a great poster to hang in the classroom! We encourage you to help us spread this important
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
- L11:592 suppresses
- L10:750 excites
- L12:517 suppresses
- L12:833 suppresses
- L8:1754 excites
- L10:203 excites
- L8:1822 suppresses
- L12:210 excites
- L10:1447 excites
- L11:1117 suppresses
- L10:920 excites
- L12:332 excites
- L12:967 suppresses
- L12:889 weakly suppresses
- L12:1232 weakly excites
- L8:1126 weakly excites unclear: scattered technical nouns
- L12:2389 weakly excites
- L10:641 weakly excites
- L11:782 weakly suppresses
- L11:431 weakly suppresses
- L12:1668 weakly excites
- L11:1074 weakly suppresses
- L12:2411 weakly excites
- L12:1136 weakly excites
Feeds later layers
- L14:2194 suppresses
- L16:669 excites
- L15:2295 suppresses
- L15:2748 suppresses
- L16:2813 suppresses
- L14:188 suppresses
- L16:2346 suppresses
- L16:1886 suppresses
- L16:2528 suppresses
- L15:1450 suppresses
- L14:924 excites
- L14:1372 excites
- L14:3028 suppresses
- L18:413 suppresses
- L14:1288 excites
- L14:121 suppresses
- L16:1225 excites
- L14:2183 suppresses
- L15:2409 suppresses
- L14:1339 excites
- L14:197 suppresses
- L18:412 suppresses
- L16:1587 excites
- L15:125 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
- ков
- ialis
- 地说
- [at
- olarity
- ności
- psychologists
- 듬
- RefCount
- strup
Suppresses
- antry
- aha
- ede
- deen
- 莺
- anceled
- edith
- ah
- الحوثي
- arbon