the modal word 'may'
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. 25 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 'may' when it means 'might be possible', as in 'there may be' or 'may occur'. When it was turned down, the text got stuck repeating 'may' or 'likely', and when it was turned up the text drifted into words like 'pain'.
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.
-
syysk. 2021|⏎|OKM-julkaisutyyppi||A1 Julkaistu artikkeli, soviteltu|<|endoftext|>Though it straddles the Arizona and Utah borders, Monument
-
regimes have short residence times, as Presidents come and go and control of Congress swings back and forth between parties. But ambitious astronomy—the kind that rewrites textbooks—may take decades to plan and execute. Building a
-
web site to request a printed copy of this publication.⏎- Our on-line publications are scanned and captured using Adobe Acrobat. During the capture process some typographical errors may occur. Please contact Sharon Hobrla, email@example.com if you notice any
-
thought that beta cells are completely wiped out (though some early research suggests there may still be some faint activity of beta cells in some people with type 1).⏎
-
are often not very precise. A more interesting question is what is the optimal type of motion.⏎I'm with you Naperlou. I would think there may be better vision models in nature than the human eye movement. The insect or bird
-
they have left their normal gathering grounds in⏎the Sea of Cortez in Baja California, Mexico. Fishermen worry that⏎a critical part of their livelihood may be gone. Many families have⏎depended on Humboldt squid since the slippery
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: may positional positional positional positional positional positional positional positional positional positional positional positional positional positional positional positional positional positional positional positional positional positional positional pos
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
12, 2013⏎Nominated early this year for recognition on the UNESCO World Heritage List, which includes such famous cultural sites as the Taj Mahal, Machu Picchu and Stonehenge, the earthen works
-
ει τετράγωνον σχῆμα τῷ λόφῳ in Einklang zu bringen, ist es unumgänglich nothwendig dass der Ausdruck Roma quadrata
-
” earthquakes, studies on active fault zones in Thailand are still minimal.⏎With all of these in mind, the Geophysics research group at Faculty of Science, Mahidol University is founded with the objective of being the “leader” in ge
-
incorrectly, an oil-analysis program is a waste of time and money. Key guidelines include:⏎- Take warmed-up machine samples while equipment is running. If that’s not possible, sample no later than 30 minutes after shutdown⏎- Use clean
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
- L23:1449 strongly suppresses 'as' in 'as ... as possible'
- L21:2006 excites "as ... as possible" phrasing
- L24:766 excites
- L21:2227 excites
- L22:1486 excites
- L22:3060 excites
- L18:2565 excites
- L20:2868 suppresses
- L19:2765 excites
- L23:1342 suppresses
- L17:1765 weakly excites
- L15:721 weakly suppresses
- L22:2775 weakly excites
- L20:2994 weakly excites
- L23:470 weakly suppresses
- L4:693 weakly excites
- L17:1279 weakly excites
- L2:180 weakly excites 'is' after focus or goal nouns
- L22:2524 weakly excites
- L20:2810 weakly suppresses
- L21:1827 weakly suppresses
- L22:1104 weakly suppresses
- L14:280 weakly excites
- L18:2247 weakly suppresses
Feeds later layers
- L26:2591 suppresses
- L27:1630 suppresses
- L27:2029 suppresses
- L27:955 suppresses
- L27:1780 weakly suppresses
- L27:2303 weakly suppresses
- L26:1677 weakly suppresses
- L27:774 weakly excites
- L27:993 weakly excites
- L27:1699 weakly excites
- L27:2642 weakly suppresses
- L26:1960 weakly suppresses
- L26:2188 weakly suppresses
- L27:1479 weakly excites
- L27:2985 weakly excites
- L26:1855 weakly excites
- L26:466 weakly excites
- L26:3005 weakly excites
- L27:2602 weakly excites
- L26:2744 weakly excites commas inside lists before 'etc'
- L26:1304 weakly suppresses
- L26:2325 weakly excites
- L27:873 weakly excites
- L26:881 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
- 怎么说
- ailable
- urable
- 습
- смотреть
- 乎
- ません
- 贯通
- いう
- 纪委监
Suppresses
- 可能
- might
- may
- 可能会
- pos
- might
- likely
- 可能是
- 可能存在
- May