'long' in 'long-term' and 'long-lasting'
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. 24 of Qwen3-0.6B. Its strongest snippets are very consistent: they mostly 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 the word 'long' when it starts a phrase like 'long-term' or 'long-lasting'. Turning it down made the model repeat 'long' over and over, so it seems tied to that word.
Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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.
-
of her last child may affect telomere length, ultimately impacting long-term health, the researchers said.⏎Longer telomeres are thought to be beneficial for long-term health, while shorter ones can signify “various chronic conditions such as cardiovascular
-
opportunity that would help sustain the museum long-term.⏎<|endoftext|>“A 200:1 ratio is very difficult to overcome
-
devices even come equipped with a back-up battery to supply instant hit once the battery dies. By not delivering nicotine, they’re not exposing one to medical risks of long-term
-
forestry and organic farming, Burung Indonesia (BirdLife in Indonesia) has worked with 27 villages around Mbeliling to improve the economic productivity and long-term viability of the local landscape. The soil is managed to avoid
-
with oxygen at the fuel cell's anode. Just as described in<|endoftext|>A new assessment finds that a majority of clean energy technologies are failing to keep pace with long-term climate,
-
be one of the largest in the area. For surrounding landowners, the project will promote different methods of long-term forest protection, such as conservation agreements, easements, and tax incentives such
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: long long long long long long long long long long long long long long long long long long long long long long long long long long long long long long long long long long long long long long long long long long long long long long long long
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
or false according to the video. This general English lesson plan finishes with a short discussion on smiling and the speaker’s ideas.⏎Because the video is about 8 minutes long, you can watch it with your students during the lesson and do the comprehension
-
a simple protein⏎. It is composed of 3 long chains wound together in a tight triple helix⏎. Each chain is 1400 amino acids⏎long. A repeated sequence of the amino acids glycine⏎make up each
-
may be secured. The home vegetable garden is worthy of<|endoftext|>MicroRNAs (miRNAs) are single-stranded, 18–24 nucleotide long, non-coding RNA molecules. They are involved in virtually every cellular process
-
under age 18 from anywhere in the world. Each contestant can submit one name, up to 16 characters long. Entries must include a short explanation and rationale for the name. Submissions
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
- L23:1490 strongly excites
- L23:2296 strongly suppresses
- L19:1712 excites
- L22:2704 excites
- L23:2238 suppresses
- L22:1122 suppresses
- L18:1641 excites the word 'long' in phrases
- L23:631 suppresses
- L10:852 excites
- L21:1628 excites
- L18:1847 excites
- L4:1978 excites
- L22:1820 suppresses
- L22:557 suppresses
- L10:1502 excites
- L18:2800 excites
- L22:2790 suppresses
- L7:1101 suppresses
- L23:2601 suppresses unclear: numbers and distance words
- L8:918 excites
- L21:658 excites
- L21:1177 excites
- L12:2377 excites
- L7:1114 suppresses
Feeds later layers
- L27:1178 strongly suppresses
- L26:1469 strongly suppresses
- L27:1469 suppresses
- L27:2280 suppresses
- L25:2963 weakly excites
- L25:2842 weakly suppresses the word a before long
- L26:1343 weakly excites
- L27:3038 weakly suppresses
- L26:2153 weakly excites
- L25:1456 weakly excites
- L25:432 weakly suppresses
- L26:2855 weakly suppresses
- L25:3037 weakly suppresses
- L26:1094 weakly suppresses
- L25:2424 weakly excites
- L27:1351 weakly excites
- L25:1808 weakly suppresses
- L25:2311 weakly excites
- L25:2823 weakly excites
- L27:1179 weakly excites
- L26:2610 weakly suppresses
- L25:1625 weakly excites
- L26:2025 weakly excites
- L26:2112 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
- ivial
- ifice
- ücü
- entious
- ṇ
- ernaut
- astery
- nex
- oge
- heit
Suppresses
- long
- long
- 长
- Long
- Long
- →long
- LONG
- longer
- (long
- .long