the 'hum' piece of words
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. 23 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 lights up on the letters 'hum' when they start a word like 'humankind', 'humbly' or 'humongous'. Turning it up made the model drift into repeated, loose phrases about relationships and people, which does not clearly tie back to the trigger.
Written by Claude from this neuron's activations and steering outputs. Confidence: medium. 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.
-
basis to abolish slavery and assure equal political rights for black Americans. As the fight goes on to assert and realize the Constitutional promise of equal representation for all, we humbly honor and celebrate the lives lost in the ongoing battle toward equality for black
-
lost its credibility”. So, today it is said there is no unified story of humankind. There is no grand narrative or what some term metanarrative.⏎
-
themselves up to their manners and population. Those living in small groups do not need the same prefect organisation as the very big populations. This thesis is valid also for humankind in small and big developed sites around the world. Capitalists and technocrats
-
any hope of avoiding catastrophic climate change means limiting CO2 concentrations below 450 parts per million by 2100. Alternative energies will only get humankind so far, and a vast removal of CO2 will be an important tool
-
to the depth of 600 and 700 feet below the sea which would mean stacking up two Statues of Liberty one over another.⏎The humungous piece of ice floating in the Antarctic has a weight of 1 trillion
-
seven million years ago. We are the only leaf remaining on our twig, and we are dependent on the same tree we share with millions of other leaves…⏎As humankind, we are all part of the same leaf…⏎The differences we have seen
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: What is the value of $ \overover{22222222222222222222222222222222222222
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
and contemporaries, including the Roman emperor Caesar Augustus, queen Cleopatra of Egypt and the Nabataean king Obodas III, issued coins emblazoned with their portraits.⏎But as the ruler of a Jewish kingdom where the
-
in a flash if you don't have a method for attacking it.<|endoftext|>304 North Cardinal St.⏎Dorchester Center, MA 02124⏎When you’re manually operating a boat, the act of operating it is
-
. Secondly, adhering to coding best practices such as modular and reusable code, efficient algorithms, and proper documentation enhances the app’s maintainability and scalability. Embracing agile development methodologies and collaboration tools promotes effective teamwork and streamlines the development process
-
They were perceived as “marginal members of the human race,” were denied the right to become naturalized U.S. citizens, and segregated to ethnic enclaves resulting in the rise of Chinatowns which were a
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
- L19:334 suppresses
- L21:1233 suppresses
- L11:2579 weakly suppresses
- L2:1923 weakly suppresses
- L17:391 weakly excites
- L16:55 weakly excites
- L20:1504 weakly excites
- L2:2278 weakly excites
- L0:169 weakly excites
- L21:2235 weakly suppresses
- L21:2981 weakly suppresses
- L1:27 weakly excites
- L21:237 weakly suppresses
- L13:792 weakly suppresses
- L5:552 weakly suppresses the word 'below' pointing to content further down
- L13:1965 weakly excites
- L13:486 weakly excites
- L20:2203 weakly suppresses
- L9:593 weakly suppresses
- L18:282 weakly excites
- L6:284 weakly suppresses
- L15:1608 weakly excites
- L7:2412 weakly suppresses
- L4:466 weakly suppresses
Feeds later layers
- L26:1908 weakly excites
- L25:1846 weakly excites
- L26:1736 weakly excites
- L25:2348 weakly excites
- L24:2717 weakly excites
- L24:2542 weakly suppresses
- L24:283 weakly suppresses
- L24:72 weakly excites
- L27:1326 weakly excites
- L25:2229 weakly excites
- L24:2317 weakly excites
- L24:1252 weakly suppresses
- L24:1682 weakly excites
- L24:741 weakly excites
- L26:2799 weakly suppresses
- L24:772 weakly excites
- L26:1811 weakly suppresses
- L27:2497 weakly suppresses
- L24:954 weakly excites
- L25:359 weakly suppresses
- L26:2332 weakly excites
- L24:800 weakly suppresses colon in clock times
- L25:2115 weakly excites
- L27:1179 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
- rence
- orrar
- rise
- érie
- 釐
- itsu
- ork
- forwards
- zier
- lane
Suppresses
- oha
- hele
- Velvet
- 势
- halten
- igan
- geme
- pecial
- pecified
- mpl