unclear: word before an event or name
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 fairly consistent: several of them 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
The snippets are mixed: it fires on words in history-style sentences, like 'in' after 'died' or 'World' before 'War', where a name or event often comes next. Turning it up made the model output the single word 'a' over and over.
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.
-
romatography<|endoftext|>President Manuel Zelaya and his opponents now in charge in Honduras remain in a standoff. Inside the country, supporters of both sides are waging mass protests, while concerns continue regarding media censorship. This crisis provides an opportunity to
-
the leading figure of the artistic school called Neo-Classicism. Born in Paris, he was deserted by his mother at age 9 after his father had been killed in a duel. Raised by uncles, he enrolled at age seventeen in the Royal
-
Indian groups throughout the Pacific Northwest, including the Nez Perce, Umatilla and Yakama.⏎Within months, hostilities broke out, now known as the Battle of Frenchtown (also known as the Battle of Wall
-
🇱 It is worth noting that the Indian contribution to victory in the First World War in the Middle Eastern area was also noted on one of the 4 stamps
-
Palm Oil Board (MPOB),⏎6 Persiaran Institusi, Bandar Baru Bangi,⏎43000 Kajang, Selangor, Malaysia.<|endoftext|>Available for download after purchase.⏎This book is designed to
-
example sentence"⏎- S: (n) fire (the event of something burning (often destructive)) "they lost everything in the fire"⏎- S: (n) fire, firing (the act of
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 way the Misle sold the information to the market, the way the Harks sold the information to the market, and the way the Misle Harks sold the information to the market. Could this be the basis for the basis of
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
.<|endoftext|>Published 4th February 2021Pre-order⏎In 1863, the US Civil War was raging. Abraham Lincoln’s Emancipation Proclamation had freed over
-
of freedom, protection and pay.⏎Additional Details from our Exhibits and Publications:⏎“And retaliation will be our practice now . . .”⏎After President Lincoln issued the Emancipation Proclamation on January 1, 1863, the
-
Ice Age that took place long ago.⏎Such an Ice Age is foretold to take place again right before the final battle:"Then will come the second Fimbulwinter. Snow will drive in from all directions; the cold will be severe
-
we were introduced to the primal ice first, and so it is when the worlds are to be destroyed, at least in a sense. This is the second Fimbulwinter, where the elements themselves begin to turn against the divine order and rip
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
- L22:2847 excites
- L21:2240 suppresses
- L22:2304 suppresses
- L20:2683 excites
- L5:2685 weakly excites
- L23:1782 weakly suppresses
- L22:2276 weakly suppresses
- L22:280 weakly suppresses
- L18:2063 weakly suppresses
- L16:1836 weakly excites
- L21:87 weakly excites
- L12:1625 weakly suppresses
- L3:736 weakly suppresses
- L23:2331 weakly excites
- L15:2353 weakly excites
- L21:1221 weakly suppresses
- L6:2808 weakly suppresses
- L22:2559 weakly suppresses
- L17:593 weakly excites
- L1:2102 weakly suppresses
- L3:1695 weakly excites opening bracket before a citation
- L2:588 weakly excites
- L15:889 weakly suppresses end of text after an abrupt cut
- L8:1330 weakly excites
Feeds later layers
- L27:231 suppresses
- L25:1586 excites
- L27:797 excites
- L27:709 excites
- L27:663 excites
- L26:1102 suppresses fibromyalgia and hour counts
- L25:1861 excites
- L25:805 weakly excites
- L27:288 weakly excites
- L26:829 weakly excites
- L25:45 weakly suppresses
- L26:484 weakly excites
- L26:213 weakly suppresses
- L26:546 weakly excites
- L26:411 weakly suppresses
- L25:1132 weakly suppresses
- L25:1314 weakly excites philosophy and ethics terms
- L25:627 weakly suppresses
- L26:1517 weakly suppresses
- L27:173 weakly suppresses
- L27:2775 weakly suppresses
- L26:2497 weakly suppresses
- L25:2924 weakly suppresses
- L25:544 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
- a
- a
- levance
- getVar
- ßer
- 颀
- adiator
- 这是一个
- 是一个
- esi
Suppresses
- arks
- sell
- ièrement
- pedest
- breakpoint
- ayas
- ords
- /cms
- Grain
- _bases