mixed: names, 'for' and 'that' after nouns
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. 21 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 pin down. It sometimes fires inside organisation names like 'International Union for Conservation' or inside a split-up name, and it goes quiet or negative on words like 'that' after a noun. Turning it up or down gave no 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.
-
61 and 1964 by biologists in Russia. Hence, it is also called as the Russian Red Data Book.⏎Currently, the International Union for Conservation of Nature maintains the Red Data Book. IUCN is the world’s
-
at Macquarie University, Sydney, Australia.⏎Browse the Lab Book to learn about the history
-
global action under the Montreal Protocol to mend the hole in the ozone layer is starting to work. In the first accurate assessment of the impact of the treaty, Macquarie University climate dynamicist Murry Salby and his colleague Lilia Desch
-
⏎Knowing which genetic variants are responsible for differences in gene activity among individuals can give powerful clues for diagnosis, prognosis and intervention of different diseases. Senior author Emmanouil Dermitzakis, Louis Jeantet Professor at UNIGE,
-
ellied.” In moments a crowd of doctors and attendants surrounded the two<|endoftext|>what is class in java⏎In Java, a class is a fundamental blueprint that defines the objects included in it.As such, it is comprised of methods, variables,
-
fruit bat (Styloctenium mindorensis) is found only on the Philippine island of Mindoro. Courtesy of ASU's International Institute for Species Exploration.⏎To date some 1.8 million species have been described,
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: it seems that a lot of people will do the same thing that does not seem to work. It will take the least amount of that which will do the best. It will take a non-mitestant, non-mitestant
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
3×2×1 is written as 8! and we call it "eight factorial." You probably have a button on your calculator labelled with [x!] that will perform these calculations for you.⏎When you shuffle a pack of cards, you
-
You merely print it out, copy the graph paper more than on to your computer, or utilize it on a electronic graph plotter. There are even programs available that permit you to print it correct out of your computer with out a printer.⏎Numer
-
the first occurrence of value in the block of memory pointed by ptr.⏎If the value is not found, the function returns a null pointer.⏎In C, this function is only declared as:⏎void * memchr ( const void *, int, size
-
building a word frequency list and using your text to visualize your data, for the most part, in an easy and simple manner.⏎Antconc: Is freeware that is cross platform for Mac, Windows, and Linux. It is a concord
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
- L19:2627 excites
- L15:2474 suppresses
- L20:1093 suppresses
- L20:2592 weakly suppresses
- L9:18 weakly suppresses text boundary after a document break
- L4:1208 weakly suppresses
- L20:96 weakly suppresses
- L17:973 weakly suppresses
- L20:1236 weakly excites
- L18:3061 weakly suppresses
- L18:1811 weakly excites 'and' linking two descriptive claims
- L12:1689 weakly excites
- L17:2489 weakly excites
- L20:1445 weakly excites
- L17:2779 weakly suppresses
- L15:1860 weakly excites
- L16:2512 weakly suppresses
- L13:135 weakly suppresses document boundary end-of-text token
- L19:2821 weakly suppresses
- L1:1872 weakly suppresses
- L18:1427 weakly excites
- L20:2778 weakly suppresses
- L19:1647 weakly excites
- L18:1151 weakly suppresses
Feeds later layers
- L22:335 suppresses
- L25:347 weakly suppresses
- L27:461 weakly suppresses
- L24:290 weakly suppresses 'to' in set phrases like 'to come'
- L23:638 weakly excites
- L24:156 weakly excites
- L23:2302 weakly suppresses
- L23:2079 weakly suppresses
- L22:1402 weakly suppresses
- L26:783 weakly suppresses unclear: word-start fragments like 'pl' and 'bl'
- L26:764 weakly suppresses
- L26:467 weakly suppresses
- L22:1171 weakly suppresses
- L22:181 weakly excites
- L25:586 weakly suppresses
- L26:246 weakly suppresses
- L25:812 weakly suppresses
- L23:1673 weakly excites
- L22:479 weakly suppresses
- L27:3015 weakly excites 'F' starting a word in a split
- L22:2958 weakly excites
- L22:402 weakly excites
- L25:2040 weakly suppresses
- L26:669 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
- peq
- loo
- će
- 鼾
- representation
- れた
- .sep
- representing
- incare
- bi
Suppresses
- returns
- computes
- ives
- >Returns
- returns
- Returns
- приним
- reads
- utton
- 体会