unclear: words mid-phrase before a fixed continuation
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 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 is hard to read. It turns up on words in the middle of a phrase, like 'their', 'two', 'mean' or 'peace' right before the rest of a set expression. It goes strongly negative on words like 'state' or 'language' that finish a noun phrase such as 'most populous state'. Turning it up or down did not produce a clear theme.
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.
-
ceptions of transcendental geometry the basis for an amusing story.⏎The very short article goes on to compare “Flatland” with “Through the Looking Glass” and their use of geometry as speculative and imaginative.⏎In trying to find more about this term
-
prosecution (sense 2)): from Outdated French, or from late Latin prosecutio(n-), from prosequi ‘pursue, accompany’ (see prosecute). Customary Application for Employment It’s our policy to adjust to all
-
like older women in the United Kingdom are using abortion like a “back-up” form of birth control even though preventing a pregnancy and killing a unique human life are two very different things. These women really need to be informed of the risks associated with
-
Antoine Pierre Joseph Marie (äNtwänˈ pyĕr zhōzĕfˈ märēˈ bärnävˈ) [key], 1761–93, French revolutionary. A
-
gods that others believe in, then we would regard them as false gods.⏎The second part of this declaration is that we believe that the Holy Prophet Muhammad, may peace and blessings of Allah be upon him, is the Messenger of Allah. Since he
-
what and how much you eat. Do you know the difference between serving and portion sizes? While the terms serving and portion often are used interchangeably, they actually mean different things.⏎A "serving" is the amount of food recommended in consumer
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 working on the course, and I have been working on the course. I have been working on the course. I have been working on the course. I have been working on the course. I have been working on the course
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
before the Great Depression brought it all to an end. Florida's economy didn't fully recover until World War II. But today, it's the most populous state in the South besides Texas, and the fourth most populous in the United States.⏎
-
result in animals with calcium and phosphorus deficiencies. In young<|endoftext|>Like most states in the Midwest, Indiana has bountiful wind resources. The Hoosier State is 13th in the U.S. in installed wind energy capacity,
-
nearly every databases – MySQL, ORACLE, PostgreSQL, Sybase, Informix to name a few.⏎What is PHP?⏎- PHP is a server-side scripting language that means its code is executed on
-
explicitly non-GMO products for their in-store brands. Because so many Ohio growers depend upon certified non-GMO seed for their livelihood, concern is growing in the state
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
- L5:1378 suppresses
- L10:2890 suppresses
- L9:590 excites
- L8:2143 suppresses
- L7:724 excites
- L7:1437 suppresses
- L9:608 excites
- L10:2920 excites
- L9:1390 excites
- L10:279 suppresses
- L8:2544 weakly suppresses
- L7:2811 weakly excites
- L10:2518 weakly excites
- L5:2566 weakly suppresses
- L7:2432 weakly excites
- L10:2094 weakly excites
- L10:1007 weakly suppresses
- L10:1631 weakly suppresses
- L10:2633 weakly suppresses
- L9:85 weakly excites
- L10:2062 weakly excites
- L10:2489 weakly suppresses
- L10:753 weakly suppresses
- L6:2082 weakly suppresses
Feeds later layers
- L12:1761 suppresses
- L13:2409 suppresses
- L12:1478 suppresses
- L15:1548 suppresses
- L15:2393 suppresses
- L12:1184 suppresses
- L14:1229 excites
- L14:2655 excites
- L13:2601 suppresses
- L13:904 suppresses
- L13:2333 excites
- L13:2538 suppresses
- L18:2141 excites
- L16:2057 excites
- L12:570 suppresses
- L12:1135 suppresses
- L14:1805 excites
- L14:2067 excites
- L13:1015 excites
- L13:2520 excites
- L13:1979 excites
- L13:2233 weakly suppresses
- L15:621 weakly excites
- L12:363 weakly 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
- Pairs
- ấu
- чат
- drivers
- ters
- Families
- Regressor
- lesc
- _pairs
- tle
Suppresses
- ITY
- short
- CED
- 已久
- amic
- 之
- .vs
- extended
- extended
- Vers