unclear: words after 'less than' or 'of'
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 fairly consistent: several of them 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 a mix of words that follow qualifiers like 'under', 'less than', 'women of', or 'developed and', such as 'five' in 'under five' and 'color' in 'women of color'. It is hard to find one clear theme. Turning it up made the text wander into repetitive, unfocused sentences without a clear topic.
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.
-
forces, a rumour starts to spread according to which NI provided the list to State A and cooperated with State A’s<|endoftext|>Find all elements less than 0 or greater than 10 and replace them with NaN⏎Getting the indices from
-
the Peace Corps. He created the Foundation Heart Ventures-Grupo de Ap<|endoftext|>is an elementary subatomic particle with infinitesimal mass⏎(less than 0.3 eV..?) and with no electric charge⏎. Neutrinos
-
on troubleshooting a refrigeration system while applying the principles of superheat and subcooling. It may also teach you the proper methods to tackle some typical troubleshooting tasks
-
Phillip Newman Specific rights apply.⏎A service of: ASD at NASA / GSFC⏎& Michigan Tech. U.<|endoftext|>Troubleshooting With Superheat, Subcooling⏎You also need to have the right tools and know-how to apply these
-
greatest challenges.⏎Roughly 1.7 million children under five years old die every year in India, says Unicef, representing more than
-
has closed by less than 20 cents. In 2010, women earned 77 cents on every dollar earned by men. Women of color fared even worse--African American women earned only 67 cents,
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 am a student and I have a question, and I am trying to write this as a blog post. I am writing this, and I have a post with a blog post, and I am writing a post summary. I am a
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
. Why are we here? What are we supposed to be doing with the time we've been granted? Achilles knew from the beginning that his years would be short, that he was destined to die at Troy, and that his legacy would forge
-
, Spain ended diplomatic relations with America and the US Navy established a blockade of Cuba, marking the official start of the Spanish-American War. The war would last less than four months, but saw the downfall of the Spanish Empire and the rise of America
-
by periods of quiescence of several months – years long. During quiescent periods, there is very little or no accretion, while during much shorter periods of activity there is an abundant accretion, due probably to disc flow
-
disparity. According to BLS data, women are paid 86 percent of the median hourly wages of men. This evaluation accounts for part-time (fewer than 35 hours) workers, which are more often<|endoftext|>Hiding information in
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
- L4:203 strongly excites
- L5:1560 strongly excites
- L5:977 excites
- L4:1315 suppresses
- L6:339 excites
- L4:1115 suppresses
- L5:2703 suppresses
- L4:1877 excites
- L6:1879 weakly suppresses
- L6:1461 weakly suppresses
- L4:661 weakly excites
- L5:559 weakly suppresses
- L5:938 weakly suppresses
- L3:2611 weakly excites
- L2:2751 weakly suppresses spiral and screw words
- L0:2023 weakly suppresses
- L4:2010 weakly excites
- L6:1129 weakly suppresses
- L4:1075 weakly excites
- L5:2328 weakly suppresses
- L4:2144 weakly suppresses
- L6:1677 weakly excites
- L3:2304 weakly suppresses
- L3:1402 weakly suppresses
Feeds later layers
- L10:2308 strongly suppresses
- L13:2869 strongly suppresses
- L11:1427 strongly suppresses
- L16:1930 suppresses
- L8:140 suppresses
- L11:1549 suppresses
- L11:2703 excites
- L11:2994 suppresses
- L13:2314 excites
- L16:1120 suppresses
- L12:455 suppresses
- L18:805 suppresses
- L8:545 suppresses
- L9:1122 suppresses
- L9:465 suppresses
- L16:1060 excites
- L11:1081 suppresses 'first' and 'inaugural' as starting points
- L12:1272 excites
- L8:1211 suppresses
- L19:2011 suppresses
- L8:2065 excites
- L13:2117 excites
- L15:1139 suppresses
- L8:1395 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
- 持久
- 久
- Demand
- ((__
- 万元以上
- ibern
- obili
- undo
- ći
- 去年
Suppresses
- 那么简单
- minor
- .createComponent
- immediate
- encers
- erture
- Ass
- aid
- ass
- minimal