unclear: pieces of longer words and analogies
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. 15 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 odd spots inside words or phrases, such as 'jigsaw', 'corollary' and 'parallel lives'. The negative side seems to show up in comparisons like 'a kind of doorman' or 'a teacher supervising'. No clear pattern, and turning it up gave unrelated maths and rambling.
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.
-
form of extractive capitalism is no different from industries that take physical resources and is reminiscent of colonial attitudes and activities.⏎Financial exclusion is a large part of the jigsaw of global poverty. But in
-
Austronesian seafarers who entered Near Oceania mixed with diverse groups already there, eventually giving birth to the distinct Lapita people, from which arose Polynesian cultures.⏎It is this sheer amount of genetic diversity that destroyed
-
life, the more you will find you use it.⏎A degree of experimentation is necessary to get exact cooking times for<|endoftext|>Here are the 15 types of distorted thinking:⏎1. Filtering: You take the negative details and magnify them,
-
taking the U.K. government to court over how the country has failed to reach its own targets over climate change.⏎Backed by crowdsourcing legal fees, Plan B’s argument that the government is discriminating towards the young by failing to make a
-
rights solely through historical precedents.⏎“Will the Court’s approach permit judges to reach the outcomes they prefer and<|endoftext|>Team first to record key event that breaks continents apart⏎(PhysOrg.com) -- Researchers have captured for the first time a geological
-
-called “resettlers” (Aussiedler and Spätaussiedler) from Eastern Europe. In this contribution, I will sketch these “parallel lives” and reflect
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 a city in the state of New York, and I have had a lot of friends, families, and communities that I have had a lot of friends, families and communities in the state of New, which I have had a
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
land instead of falling down into the fires beneath where everything was burned. Smith's Bible Dictionary gives a description of this valley — which was something like one of our city dumps today where trash and rubbish are burned up.⏎If anything, especially a dead
-
the input-output pair. It is called supervised learning because the process of an algorithm learning from the training dataset can be thought of as a teacher supervising the learning process.⏎Some use cases that are based on supervised machine learning
-
‘wrong’ kinds of carbs contribute to it.⏎Insulin resistance occurs when the body stops responding to this critical hormone. Think of insulin as a kind of doorman. Normally, it triggers a reaction that allows glucose to pass through cell walls,
-
Fraud⏎Mail fraud describes any action meant to defraud somebody that makes use of the United States Postal Service or any other interstate mail carrier. This would include any action 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
- L14:969 excites
- L12:2690 suppresses
- L12:2520 suppresses
- L14:1275 suppresses
- L12:890 excites
- L11:67 excites
- L13:2037 weakly excites
- L13:1714 weakly excites
- L7:204 weakly suppresses
- L13:2062 weakly excites
- L13:142 weakly excites
- L10:2079 weakly excites
- L12:1152 weakly suppresses
- L14:1872 weakly suppresses
- L14:1539 weakly suppresses
- L5:2278 weakly excites
- L13:2461 weakly excites
- L8:1878 weakly suppresses
- L13:1344 weakly excites
- L12:2570 weakly suppresses
- L14:145 weakly suppresses
- L11:2403 weakly excites
- L10:1849 weakly excites
- L13:785 weakly suppresses
Feeds later layers
- L19:443 excites
- L19:746 excites
- L18:336 excites
- L19:313 suppresses
- L20:2189 suppresses
- L16:3008 excites
- L22:177 excites
- L20:409 excites
- L19:253 excites
- L26:140 suppresses
- L20:253 excites
- L20:657 excites
- L16:1403 excites
- L21:372 suppresses
- L22:383 suppresses
- L21:3052 suppresses
- L20:432 excites
- L20:1810 suppresses
- L26:101 suppresses
- L19:428 excites
- L21:2752 suppresses
- L21:401 excites
- L20:893 suppresses
- L18:839 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
- CellValue
- Substring
- ecimal
- Multiply
- sin
- 痧
- (sin
- adecimal
- cosine
- sin
Suppresses
- 聘请
- bankrupt
- personnel
- headquarters
- staff
- year
- rented
- outreach
- 国家级
- 年度