'Pr' starting capitalised names
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. 26 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
It fires on the 'Pr' at the start of a capitalised name, most clearly in 'Pravda', 'Prussian', 'Prinz' and 'Prudhoe Bay'. Turning it up did not make the model write anything about these names, and the outputs wandered into unrelated text.
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.
-
and bread; All power to the soviets.⏎He was caught up in a struggle in the Bolshevik party and called a betrayer of socialism because of the Pravda publishing support for the war effort and the provisional gouvernment.⏎He
-
ast, Russia), also called "the scattered church of Prussian Lithuania." The Mennonites living here had much to endure from the dragoons of the Prussian King Friedrich Wilhelm I, who finally expelled them from the district. The [[
-
ideal platform for this work.⏎Angela Prinz from Germany is the team leader. She confirms that Eunice had a
-
places of Polish Prussia.⏎The Dutch Committee of Foreign Needs organized a general offering in May 1
-
parts and the diagrams found in a computer magazine. It was Wennberg who came up with the man to do just that. After discussion with another friend, Wes Prichard, Prichard suggested to Wennberg<|endoftext|>Home > Preview⏎The
-
overthrow of the Romanov Dynasty⏎Role of Stalin in the 1917 Revolution⏎He was released from exile and became part of the editorial board of Pravda.⏎He was elected to the
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: How can an untrained staff member which seenoach seenoach seenoach seenoach seenoach seenoach seenoach seenoach seenoach seenoach seenoach seenoach seenoach se
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
cers, acid reflux, and gastritis. The mucilage of the plant’s pulp has restorative and protective qualities to the stomach lining, and applied externally to sprains, burns, and bruises reduce swelling, discoloration and inflammation while speeding
-
maintain healthy blood sugar levels. Ginger, another key spice in pumpkin pie, is well-known for its benefits as a digestive aid and is particularly helpful for nausea. Cloves can benefit digestion and the immune system, and nutmeg can
-
of yellowfin tuna (Thunnus albacares) in nearshore areas in French Polynesia, determined using ultrasonic telemetry⏎Centre IRD de Montpellier, BP 5045, 34032 Mont
-
RNA from chemically complex plants⏎More information: The draft genome of the carcinogenic human liver fluke Clonorchis sinensis, Xiaoyun Wang, Wenjun Chen, Yan
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
- L25:741 strongly excites
- L22:390 excites
- L24:1707 suppresses
- L24:2239 excites
- L23:493 excites
- L24:25 suppresses
- L20:692 suppresses
- L18:2989 weakly excites
- L18:345 weakly excites
- L24:242 weakly excites
- L24:1421 weakly suppresses
- L20:2494 weakly excites
- L7:122 weakly excites
- L8:7 weakly excites
- L13:2993 weakly excites
- L18:569 weakly excites
- L9:2440 weakly excites
- L4:784 weakly suppresses
- L19:2358 weakly excites "any of" before "these" or "the"
- L19:353 weakly excites
- L12:2167 weakly suppresses
- L15:2621 weakly suppresses
- L20:2581 weakly excites
- L12:469 weakly excites
Feeds later layers
- L27:2782 suppresses
- L27:467 suppresses
- L27:713 suppresses
- L27:239 excites
- L27:587 weakly excites
- L27:1266 weakly excites
- L27:1807 weakly suppresses
- L27:1584 weakly suppresses
- L27:2957 weakly excites
- L27:46 weakly excites
- L27:2841 weakly suppresses
- L27:2988 weakly suppresses capital M starting a word
- L27:988 weakly excites word before 'needed' or 'required'
- L27:2034 weakly suppresses
- L27:1155 weakly suppresses just before a pronoun for God
- L27:977 weakly excites
- L27:2111 weakly excites
- L27:1947 weakly suppresses
- L27:72 weakly excites
- L27:775 weakly suppresses
- L27:1644 weakly excites
- L27:1757 weakly excites
- L27:1116 weakly excites
- L27:1880 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
- };⏎⏎⏎
- зон
- '];?></
- };⏎⏎⏎
- };⏎⏎⏎⏎
- };⏎
- '};⏎
- }; ⏎ ⏎
- };⏎⏎
- };⏎
Suppresses
- ())
- ) ⏎
- ")
- 另一半
- ()>⏎
- ")
- ") ⏎
- )⏎
- &)
- "]