middle pieces of split surnames
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. 22 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 fires on a fragment in the middle of a person's surname or a name that the tokenizer has split into pieces, such as the 'L' in Lerdorf or Llosa. Turning it down made the model get stuck repeating the word 'with', while turning it up gave loose, unrelated sentences.
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.
-
the server.⏎- PHP was created by Rasmus Lerdorf in 1995.⏎- Though PHP was originally known as
-
bad soil” or “land without anything growing on it”.⏎Firstly, peculiarity of this place noticed in 1936 by V.Vahrameev and later in
-
been accused of being traitors because they left their homelands, from Dante to Joseph Conrad and Joyce, García Marquez and Gunther Grass and Vargas Llosa? And whatever the explanation that those posing the question—
-
writes, but rather evaluate him by where he lives, or by the location of "the room" from which he writes, as the Peruvian writer Vargas Llosa noted in his commentary on the issue in one of his articles.⏎This
-
support the case for a largely periodic, coordinated, and intermittently catastrophic geologic record, which is a departure from the views<|endoftext|>Dr. Folkert van Werven⏎Dr. Alexi Goranov⏎Cells, regardless of whether they
-
al scientists with a strong interest in teaching.⏎Folkert van Werven, and Alexi Goranov. 7.342 To
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: with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with with
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
the sensation of vertigo. The condition occurs most often in people 50 and older and is more common in women than in men.⏎Doctors treat BPPV by
-
range often experiences greater exposure to bacteria (via school), leading to infection and tonsillitis. However, adults can also experience chronic tonsillitis which can be treated by receiving a tonsillectomy.⏎The surgical procedure for
-
sensors, buttons, alligator clip pads and more. Build projects with Circuit Playground in a few minutes with the drag-and-drop MakeCode programming site, learn computer science using the CS Discoveries class on code.org, jump into CircuitPython to learn
-
. The Cyberknife technology works so precisely that the administered dose only reaches the tumor and spares surrounding tissue, organs or nerves.⏎Which tumors and lesions can be treated with Cyberknife?⏎The Cyberknife System is designed for radiosurgical treatment of tumors
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
- L19:2736 strongly excites
- L21:523 strongly excites
- L16:2312 excites
- L18:2599 suppresses
- L19:1689 excites
- L18:2460 excites
- L14:1330 excites
- L19:2321 suppresses
- L20:2484 suppresses
- L19:2940 excites
- L21:1488 suppresses
- L20:2516 suppresses
- L21:2223 suppresses "to" after "used" (purpose)
- L20:3044 excites
- L21:2595 excites
- L20:501 suppresses
- L20:1525 excites
- L18:1127 excites
- L18:2975 excites
- L18:2903 excites
- L10:430 suppresses
- L17:2474 suppresses
- L20:1758 excites
- L16:2321 suppresses
Feeds later layers
- L26:741 strongly suppresses
- L24:921 strongly suppresses
- L27:603 strongly suppresses
- L24:883 strongly suppresses
- L26:223 excites
- L23:984 suppresses
- L24:2419 excites
- L27:1443 suppresses
- L23:684 weakly excites
- L27:2153 weakly suppresses
- L23:1235 weakly suppresses
- L27:1606 weakly suppresses
- L27:1323 weakly suppresses
- L25:2139 weakly excites capital N starting an abbreviation
- L23:2461 weakly suppresses
- L24:2530 weakly excites
- L23:1292 weakly suppresses
- L25:1114 weakly suppresses
- L27:430 weakly suppresses
- L27:2572 weakly suppresses
- L23:1073 weakly excites
- L23:2303 weakly excites
- L24:1741 weakly excites
- L23:605 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
- izu
- osterone
- 一字
- edef
- ]|[
- ethyst
- achen
- ','=',$
- utherford
- ackle
Suppresses
- with
- with
- With
- With
- _with
- →with
- WITH
- với
- avec
- .with