"that" introducing a clause or "that is"
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. 23 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 fires on the word "that" after a claim, especially in phrases like "positive rights; that is" or "prove that" or "believe that", where a new statement is about to start. Turning it up made the model repeat "the" over and over, so the steering results are not very informative.
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.
-
unitarians seek to bolster social capital and the institutions of civil society. . . .⏎Central to many communitarians' philosophy is the concept of positive rights; that is, rights or guarantees to certain things. These may include free education, affordable
-
home, or stay at his, it only takes one security incident to cast doubt over the safety of this model. The company created a way to check and prove that users are “real people.”⏎And this is where the Digital Identity story begins.⏎
-
, and laws limiting such things as pollution and gun violence.⏎A common objection is that by providing such rights, they are violating the negative rights of the citizens; that is, rights to have something not done to you. For example, taking money
-
“States should identify marine ecosystems exhibiting high levels of biodiversity and productivity and other critical habitat areas and should provide necessary limitations on use of these areas, …(and that) Priority should be accorded to seagrass beds…”;⏎Reminded by
-
responsible for the vast majority of federal death sentences. By addressing the unique geography, we identify a possible explanation for the racial distortions in the federal death penalty: that federal death sentences are sought disproportionately where the expansion of the venire from the county
-
even after filtering through Chamberland filter candles⏎, in 1892⏎. However, he remained convinced despite repeated failures to produce evidence, that bacteria were the infectious agents. In 1898, Martinus Be
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: Please be very very specific about who and who be back to be back at who be back at who. Please be very very specific. Please be very very very specific. Please be very very very very very very very very very very very very
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
20|⏎|Order Information:||Web: http://www.tandfonline.com/pricing/journal/VECE20|⏎When requesting a correction, please mention this item's handle: RePEc:taf:jeduce
-
a woman's life to a contemporary meditation on the meaning of the story of Lot's wife.⏎This article is available to subscribers only. Please subscribe for full access—subscriptions begin at $2.95. Already have an online account? Log
-
MENA region participated in the workforce. Today, that number is 25 percent—still half the rate of the global average.1⏎Gayle Tzemach Lemmon, a senior fellow at the Council on Foreign Relations as well as
-
precautions to help control the spread of the flu and norovirus in their facilities.⏎Kim LaFreniere, Ph.D., associate research fellow at the Clorox Professional Products Company, offers cleaning professionals four infection control tips, provided below.⏎
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
- L17:1282 excites
- L19:1100 suppresses
- L17:2684 suppresses
- L20:2430 excites
- L16:1303 excites 'that' starting a clause after a noun
- L3:1003 excites
- L20:982 excites
- L18:1389 weakly excites
- L0:684 weakly suppresses
- L21:191 weakly suppresses
- L20:1832 weakly excites 'as' introducing a reason
- L7:1098 weakly excites
- L18:645 weakly suppresses
- L17:253 weakly excites
- L18:598 weakly suppresses
- L6:2313 weakly excites
- L19:1972 weakly suppresses
- L18:2031 weakly suppresses
- L15:3031 weakly suppresses
- L21:172 weakly suppresses
- L7:313 weakly excites
- L17:2695 weakly suppresses
- L0:442 weakly suppresses
- L6:1646 weakly suppresses
Feeds later layers
- L27:1042 weakly excites
- L27:149 weakly suppresses
- L27:796 weakly suppresses
- L24:298 weakly suppresses 'would' in hypothetical or past statements
- L25:247 weakly excites
- L24:2151 weakly suppresses
- L26:1092 weakly suppresses
- L26:368 weakly suppresses
- L27:478 weakly excites
- L27:2250 weakly suppresses
- L26:1007 weakly excites
- L25:2123 weakly excites
- L26:2984 weakly suppresses
- L27:108 weakly suppresses
- L26:2385 weakly suppresses
- L24:306 weakly excites
- L25:1404 weakly suppresses
- L26:2490 weakly excites
- L26:693 weakly excites
- L26:127 weakly suppresses 'co' starting a word piece
- L24:261 weakly excites
- L27:2972 weakly suppresses
- L26:1891 weakly suppresses
- L27:121 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
- a
- those
- no
- laughter
- éné
- Prediction
- none
- а
- ging
- af
Suppresses
- 過來
- депут
- 有多大
- ạp
- ');</
- 로
- pline
- ожно
- ---</
- 请联系