commas inside lists of items
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. 2 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 responds to commas that separate items in a list, such as 'red, green, and blue' or 'lineups, show-ups, and photographic lineups'. Turning it up made the text wander into unrelated topics, so the list habit did not show up clearly in the output.
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.
-
, change the -f to -v and add -es⏎In compound words, make the main word plural.<|endoftext|>Mexico has three classifications of holidays: statutory,
-
: At the completion of this unit the⏎Chapter Fifteen Summary⏎The police frequently use three procedures to help by witnesses identify criminals: lineups, show-up, and photographic lineups. Lineups involve showing the witness a group of possible suspects
-
green and blue flash. Not a myth even in a land of runes, the colorful but elusive glints are caused by atmospheric refraction enhanced by long, low, sight lines and strong atmospheric temperature gradients.⏎Authors & editors
-
strikes. And they can dazzle the world with their cute googly eyes.⏎The world looks colorful to our eyes with only three color receptors — red, green, and blue. But for a mantis shrimp, a small marine crustacean
-
especially topics that allow students to explore possible majors in communication, business (e.g., marketing, finance), the social sciences (e.g., political science, sociology, global studies), writing, and more. The course will inspire students to think critically
-
that is, the tow-plane which pulls them into the air.⏎Within a piston engine, the pistons can be arranged in four ways: radial, in-line, oppositional and "V." The radial engine has pistons arranged in a circle
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 thing is a thing and the thing is a thing and the thing is a thing and the thing is a thing and the thing is a thing and the thing is a thing and the thing is a thing
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
text in Microsoft Word files.⏎- Heap variables that are freed with free()⏎- Automatic variables left on the stack of languages like C or garbage collected in languages like Java.⏎Byers, Simon. Scalable Exploitation of, and Responses to
-
. It provides the means to create engaging user interfaces, implement complex algorithms, and integrate various features seamlessly. By leveraging programming languages such as Java, Swift, or Kotlin, developers can build apps for different platforms like iOS, Android, or cross-platform
-
domain names, such as mail services.⏎Why is DNS important?⏎Domain names are fundamental to the Internet. Every Internet-connected device, whether a personal computer, smart phone, or gaming console, looks up each name in the global DNS, and uses
-
States with a key source of the high-quality lithium needed to produce lithium-ion batteries used in a range of digital devices and green technologies, from computer tablets and cell phones to electric cars and microgrids.⏎MORE: Salton Sea coverage⏎The
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
- L1:641 excites
- L1:698 excites
- L1:647 weakly excites
- L1:17 weakly suppresses
- L1:307 weakly excites
- L0:539 weakly suppresses
- L1:436 weakly excites
- L0:149 weakly excites
- L1:1534 weakly suppresses
- L1:414 weakly excites programming languages and software names
- L1:689 weakly excites
- L0:159 weakly suppresses
- L1:381 weakly suppresses
- L1:2568 weakly excites
- L0:301 weakly suppresses
- L1:3030 weakly suppresses
- L1:853 weakly excites
- L0:184 weakly excites
- L1:582 weakly suppresses
- L1:442 weakly suppresses
- L0:1199 weakly excites
- L0:638 weakly suppresses
- L1:443 weakly suppresses
- L1:272 weakly excites
Feeds later layers
- L4:100 strongly excites
- L6:499 excites
- L6:2117 excites
- L5:63 excites
- L6:123 excites
- L3:404 excites
- L7:33 suppresses
- L7:664 excites
- L4:348 suppresses
- L6:497 suppresses
- L5:477 excites
- L3:757 suppresses
- L5:42 excites
- L5:693 excites
- L6:280 excites
- L3:666 excites
- L5:310 suppresses
- L7:120 suppresses
- L3:376 excites
- L4:508 suppresses
- L7:51 excites
- L6:111 excites
- L9:600 excites
- L7:80 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
- ')</
- anches
- repr
- HomeController
- azeera
- .cs
- angepicker
- 瘙
- ữa
- imately
Suppresses
- 渎
- ึ
- dign
- 現在
- aland
- lege
- INE
- mie
- ocate
- IDA