'the' after 'like' or 'stay'
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 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 tends to fire on the word 'the' when it comes after 'like' in comparisons, such as 'like the grocery store'. Its strongest negative cases involve 'and' between superlatives. Turning it up gave repetitive, wandering text.
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.
-
hundreds in a puritan society. In his novel,The Scarlet Letter, Nath<|endoftext|>What they learned after only seven years is worth remembering (or learning, as the case may be).⏎The United States has operated under two constitutions. The first
-
large telescope or sending a spacecraft to another planet, for example, requires policymakers to embrace a common goal and to stay the course for many years, shielding scientific workers from the shifting whims of politicians.⏎The
-
Cooper.⏎Buy the hardcover edition of⏎I Heard the Willow Weep for $6.00. (Retail price: $1
-
But now that she is a reader, surround her with print. Get a newspaper delivered. Get her a library card and make the library a regular stop, like the grocery store and the dry cleaner. And get over your view of what "proper
-
landed on the “dark side” of the moon! The “dark side” … like the evil side? Nah—it’s actually the South Pole-Aitken Basin area of
-
One of the Greatest Forces⏎The power of a successfully communicated thought, from one human mind to another, is one of the greatest forces we know. But like the tango, it takes two to communicate
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: the most commonestestestestestestestestestestestestestestestestestestestestestestestestestestestestestestestestestestestestestestestestestestestestest
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Amazon.com.<|endoftext|>The world’s seabeds are littered with wrecks and treasures. Despite many decades of increasingly active searches for the best, the rarest and the most valuable of them
-
class on code.org, jump into CircuitPython to learn Python and hardware together, TinyGO, or even use the Arduino IDE. Circuit Playground Express is the newest and best Circuit
-
the party has not completely abandoned state planning and nationalisation. But less has changed politically. The party - based on the Soviet model - runs the country in much the same way it has always done.⏎What about the future?⏎Some believe the opening
-
writing career, Vladimir Nabokov was frustrated by both the Freudian and behavorialist schools of psychology that reigned in his day, seeking a richer and more complex view of the mind than either provided. In a long, incis
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
- L5:2787 strongly excites 'and' in 'more and more' pairs
- L4:1727 weakly excites
- L2:473 weakly excites
- L4:682 weakly suppresses
- L4:344 weakly excites
- L2:2182 weakly excites
- L2:70 weakly excites
- L5:158 weakly suppresses
- L1:2801 weakly excites
- L6:16 weakly excites
- L5:2844 weakly excites
- L4:801 weakly suppresses
- L0:2842 weakly suppresses
- L0:960 weakly excites
- L5:225 weakly excites
- L6:1004 weakly suppresses
- L2:3065 weakly excites
- L5:2345 weakly excites
- L3:2071 weakly excites
- L1:2032 weakly excites
- L4:2270 weakly excites
- L1:1809 weakly suppresses
- L6:107 weakly suppresses 'organisms' and 'Code' in reference contexts
- L6:1542 weakly excites
Feeds later layers
- L10:94 strongly suppresses
- L19:395 suppresses
- L17:446 suppresses
- L8:423 suppresses
- L16:286 suppresses
- L8:71 suppresses
- L22:969 suppresses
- L16:592 suppresses
- L15:2378 suppresses
- L24:118 suppresses
- L16:1231 suppresses
- L16:283 suppresses
- L9:929 suppresses 'each' in 'each other'
- L21:327 suppresses
- L16:415 suppresses
- L13:74 suppresses
- L13:2580 suppresses
- L16:582 suppresses
- L14:713 suppresses
- L8:1032 suppresses
- L15:1470 suppresses
- L27:2557 suppresses
- L25:2986 suppresses '1' in 21st or 11th, before 'st'
- L20:944 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
- heels
- SPDX
- voi
- però
- vu
- vero
- HostException
- võ
- nik
- уш
Suppresses
- est
- 最快
- er
- 了一个
- 级以上
- 更多
- more
- ar
- s
- ingly