capital T starting a name
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. 25 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 lights up on a capital T that begins a word or name, such as Torchwood, Tijuana or Tropicana, especially when the word has been split into pieces. Turning it up made the model's writing loop and drift without a 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.
-
Being undead, Owen Harper (Torchwood) has no need for oxygen, allowing him to survive underwater.⏎Surv
-
visits, secret sign, skilled carvers, St Bosa of York, St William of York, stained glass, stonemasonry, stonework, Tadcaster, Vale of York, weathering
-
and the Ukrainian Cossacks (козаки), respectively. Somewhat less well known are the Polish Cossacks (Kozacy) and the Tatar Cossacks (Nağaybäklär). The name “C
-
19⏎Shohei Doi,Kazunori Inamasu, Shoko Kohama, Atsushi Tago⏎Tetsuro Kobayashi, Atsushi Tago, Kyu S.
-
production stage manager for numerous Broadway and regional productions, as well as New York City Opera. His diverse career in live performance presentations includes: Entertainment Operations Manager for the Tropicana, Production Coordinator for Lincoln Center Festivals, and Associate Producer/Writer
-
idji High School and students from Schoolcraft to discuss their concerns about bullying and ways in which they think it could be prevented.⏎Students like Thomas Caddy and Tia Siddens, 9th graders at Bemidji High School
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 need to have a live live live live live live live live live live live live live live live live live live live live live live live live live live live live live live live live live live live live live live live live live live live
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
, heals wounds and improves skin strength.⏎Hesperidin and pect<|endoftext|>Bacterial whole-cell biosensors were used to measure the concentration of chlortetracycline (CTC) in the feces of pigs. In this
-
Whether commercially or domestically prepared, the principles behind baking powder formulations remain the same. The acid-base reaction can be generically represented as shown: NaHCO3 + H+ ? Na+ +
-
ographic method for the direct determination of ammonium, potassium, and sodium in geologic materials is described. Samples are decomposed with a mixture of hydrofluoric and hydrochloric acids in a sealed polycarbonate bottle heated in a microwave oven
-
infarctions ("heart attacks") in the past⏎- Individuals who have congestive heart failure of any cause⏎- Individuals with intrinsic heart muscle disease such as hypertrophic cardiomyopathy, idiopathic dilated cardiomyopathy, and arrhythm
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
- L22:316 excites
- L24:1220 excites
- L23:1427 weakly suppresses
- L20:1040 weakly suppresses
- L7:2724 weakly suppresses
- L23:2842 weakly suppresses
- L21:1112 weakly suppresses
- L20:238 weakly excites
- L23:724 weakly suppresses
- L1:87 weakly suppresses
- L12:1729 weakly excites
- L7:2341 weakly suppresses
- L19:529 weakly excites
- L23:2489 weakly suppresses
- L4:2220 weakly excites
- L2:1382 weakly excites
- L8:2572 weakly excites
- L3:564 weakly suppresses
- L5:689 weakly suppresses
- L11:2738 weakly excites
- L18:408 weakly excites
- L2:515 weakly suppresses
- L8:950 weakly suppresses
- L3:461 weakly excites
Feeds later layers
- L26:1474 weakly suppresses
- L26:1346 weakly excites
- L27:1568 weakly suppresses
- L27:1355 weakly suppresses
- L26:1533 weakly excites
- L26:2860 weakly suppresses
- L26:2669 weakly excites
- L27:880 weakly excites
- L26:2572 weakly excites
- L27:1529 weakly excites
- L27:2458 weakly suppresses
- L26:1434 weakly suppresses
- L27:1654 weakly excites
- L26:1366 weakly suppresses
- L26:984 weakly excites
- L26:2853 weakly excites
- L26:2620 weakly suppresses
- L27:2213 weakly excites
- L27:686 weakly excites
- L26:1724 weakly excites
- L27:2502 weakly suppresses
- L26:2729 weakly excites
- L26:1100 weakly suppresses
- L27:2534 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
- ỵ
- ibile
- BarButtonItem
- だが
- (optarg
- 和发展
- tà
- だけど
- komen
- hibit
Suppresses
- ric
- concent
- gee
- madness
- 泄
- 甾
- art
- extends
- processor
- RIC