second halves of split names and acronyms
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. 21 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 the later pieces of a name or abbreviation that the tokenizer has split into parts, such as the 'lu' in RLUIPA or the 'M' in MPA. Turning it up led to unrelated text, so the steering results do not add much.
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.
-
houses of worship, and religious schools. The full text of RLUIPA is available at http://www.usdoj.gov/crt/housing/housing_rluipa.htm. Below is a summary of the law’s key provisions, with illustrations
-
religion?⏎Íocaíocht & Slándáil⏎Déantar do chuid sonraí íocaíochta a phróiseáil go slán. Ní dhéanann muid sonraí cártaí cre
-
known as Kurd Dagh (Çiyayê Kurd in Kurdish), meaning Mountain of the Kurds. The major urban center is Efrîn (‘Afrin in Arabic), with an urban population of approximately 80,00
-
distribution and host preference of lianas in a tropical dry evergreen forest in India.⏎Vivek, P., & Parthasarathy, N.⏎Lianas are structurally dependent life-form that rely mostly on trees to access the forest
-
recorded with various hanja phonetically approximating its native Korean name: 斯盧 (사로, Saro), 斯羅 (사라, Sara), 徐那(伐) (서나[벌], Seona
-
and skill level different and above that for⏎passenger cars, particularly when the vans are fully loaded or drivers⏎experience an emergency situation.”3 While the NTSB had⏎previously made recommendations about driver training and use policies⏎for fifteen
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 to the present in the way of the real, the way of the real that is the by the way of the real, the by the way of the real, by the way of the real, the by the way of the
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
electronic forms of online payment.⏎9. Be on guard against scams hiding behind online coupon offers.⏎10. Be careful if you download banking software onto
-
but their design limitations make them of little value for risk evaluation⏎- Assessing whether fluoride constitutes a risk factor for osteosarcoma is complicated by the rarity of the disease and the difficulty of characterizing biologic dose because of the ubiquity
-
ism was shown off at the 1933 Radio Exhibition at the Olympia exhibition space in London. The most intriguing aspect of its design is the world map sitting just below the screen. One
-
effort to prove we could create a commercial quality game using NASA content that is fun. NASA was looking for a project at the same time the Army Game Studio was looking to branch out from America’s Army. It helped that Army Game Studio and NASA
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
- L18:569 suppresses
- L0:1224 excites
- L6:13 suppresses
- L0:1301 excites
- L8:7 suppresses
- L0:178 excites
- L8:10 suppresses
- L2:75 excites
- L7:110 excites
- L1:1464 suppresses
- L6:7 excites
- L15:82 excites
- L13:983 weakly suppresses
- L5:46 weakly suppresses
- L16:2176 weakly excites
- L11:322 weakly excites
- L5:50 weakly suppresses
- L5:169 weakly suppresses
- L3:56 weakly suppresses
- L4:97 weakly excites
- L11:456 weakly excites
- L5:110 weakly excites
- L16:1392 weakly excites
- L8:4 weakly suppresses
Feeds later layers
- L27:188 excites
- L27:206 excites
- L26:214 excites
- L22:80 excites
- L22:85 excites
- L22:84 excites
- L26:242 excites
- L22:99 excites
- L22:79 excites
- L22:66 excites
- L22:982 suppresses
- L22:128 excites
- L22:112 excites
- L23:44 excites
- L24:1017 suppresses
- L22:92 excites
- L22:58 weakly excites
- L22:269 weakly excites
- L22:88 weakly excites
- L22:90 weakly excites
- L22:98 weakly excites
- L22:59 weakly excites
- L23:666 weakly suppresses
- L22:337 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
- @"⏎
- UserController
- ToProps
- encodeURIComponent
- 忘れ
- eventType
- URLConnection
- strawberries
- Controllers
- addObject
Suppresses
- ucle
- andal
- ktor
- IFE
- ine
- лиз
- orado
- bia
- eload
- лен