opening bracket before an acronym
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 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
This neuron fires on the opening bracket that comes right after a full name, just before its short form, such as the bracket before "ACA" or "AAP". It goes quiet, and sometimes slightly negative, on opening quote marks. Turning it down made the text fill with blank placeholders, and turning it up gave ordinary rambling, so steering shows no clear theme.
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.
-
US citizenship were also three times less likely to have private or public health insurance, resulting in minimal to no access to comprehensive reproductive health care. The Affordable Care Act (ACA) preserved a 1996 restriction requiring non-citizens to
-
choice of format and writing style.⏎English 102. It is assumed that you already have basic knowledge and skill in critical thinking, library research, MLA (Modern Language Association) documentation, American English grammar, style, usage, and persuasive
-
published books and computational analyses of social media data - to disentangle the dynamic relationship between Japanese public opinion and Japan–South Korea relations.⏎
-
in the individual market. These plans can be purchased directly from insurance providers or through state-based exchanges, such as the Health Insurance Marketplace established under the Affordable Care Act (ACA).⏎c. Government Programs: The US government provides medical insurance through programs like
-
shift in patient care. Advances in endoscopic techniques and in imaging help achieve these goals. Among endoscopy techniques are:⏎- Endobronchial ultrasound (EBUS)⏎- Electromagnetic navigation (EMN) with volumetric high
-
32K code size limit; you can break up your code into a main portion (of at most 32K), and a number of libraries (each of at most 32K).⏎GLib (pronounced ``gee
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: __ __ __ __ __ __ __ __ __ __ __ __ __ __ __ __ __ __ __ __ __ __ __ __ __ __ __ __ __ __ __ __ __ __ __ __ __ __ __ __ __ __ __ __ __ __ __ __
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
sharp phrases, interesting metaphors, rhyming and code-switching. Calò words and phrases, such as “la neta,” (the truth) “trucha,” (vigilant)
-
lion had an important role in Roman mythology, since it was associated with the god Hercules. Notoriously, lions were integral to the form of capital punishment known as ‘damnatio ad bestias’, whereby condemned criminals were pitted against
-
Denudate [dih-NOO-deyt ](verb): to make bare; strip; denude.⏎This fun word comes from the latin word “denudatus” which means “naked” or “ bare.”⏎With the
-
auxilliary forces rushed to the rescue of the farmer, o'erwhelming the⏎enemy and with one fell swoop, bringing on them consternation and⏎ruin dire. It appears that the foe, or their progenitors,⏎had
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
- L25:239 strongly excites
- L20:563 excites
- L19:509 excites
- L22:65 suppresses
- L21:321 excites
- L21:29 excites
- L19:124 suppresses
- L24:77 excites opening parenthesis after a name
- L23:80 excites blank slot before a year or number
- L19:1207 suppresses
- L20:1585 excites
- L25:222 suppresses
- L19:1174 excites
- L18:76 suppresses blank spot before a number
- L23:90 suppresses
- L14:140 suppresses
- L20:433 suppresses
- L23:262 suppresses
- L16:1421 suppresses
- L20:499 excites
- L24:141 suppresses
- L23:864 suppresses
- L24:94 excites
- L25:17 excites
Feeds later layers
- L27:921 suppresses
- L27:135 weakly suppresses
- L27:527 weakly suppresses
- L27:802 weakly suppresses
- L27:2890 weakly excites
- L27:580 weakly excites
- L27:526 weakly suppresses
- L27:476 weakly suppresses
- L27:514 weakly suppresses
- L27:912 weakly suppresses
- L27:553 weakly suppresses
- L27:898 weakly excites
- L27:2035 weakly excites
- L27:186 weakly suppresses
- L27:450 weakly excites
- L27:573 weakly suppresses
- L27:213 weakly excites
- L27:337 weakly suppresses
- L27:327 weakly excites
- L27:722 weakly suppresses
- L27:485 weakly suppresses
- L27:561 weakly suppresses
- L27:1982 weakly excites
- L27:458 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
- 촌
- 一期
- 两岸
- tile
- 化合
- 嫉
- -feature
- generation
- donn
- 省政府
Suppresses
- ing
- ging
- rys
- ming
- ries
- ister
- azer
- ges
- nelle
- медицинск