'that' after 'so ... that'
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. 3 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' in phrases like 'so inflexible that' or 'so intense that', where a result clause is about to start. Turning it up did not visibly make the model write this kind of sentence, so the test results don't confirm it.
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.
-
. Some come to the centre hungry (it provides meals) or speaking no English. Others arrive with behaviour problems. Parents’ work schedules are often so inflexible that Federal Hill must cover basic logistics beyond school pick-up and drop-off. Ms Fernandez
-
engineering, and mathematics. The NOVA Awards program combines knowledge of STEM concepts with hands on activities.<|endoftext|>The sheer amount of gold in Fort Knox was so great that Auric Goldfinger, the villain of the 1964 James Bond
-
, so common in most galaxies, has been stripped away, leaving behind a dark matter shadow. These intriguing galaxiesknown as dwarf spheroidalsare so faint that, although researchers believe they exist throughout the universe, only those relatively close to Earth
-
vey & Dr. Sigwart can diagnose and screen for a host of maladies thanks to sophisticated imaging technology. X-rays are now so frequently used and so effective that most of us take their capabilities for granted—but it was not so long ago a
-
Chip teaches them how to safely use the internet.⏎The BSA feels so strongly about internet safety that the Cub Scouts must complete the Cyber Chip in order to earn their rank badge.
-
Rice was found its way into Japan about 2000 years ago, via the Asian continent. However, the Japanese climate was so hot and humid that the rice tended to become mouldy very quickly. The people, ever eager to
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: I have been thinking about lately: I have been thinking about lately: I have been thinking about lately: I have been thinking about lately: I have been thinking about lately: I have been thinking about lately: 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.
-
save 5% of my operating costs”?) so we need to define requirements at some intermediate level in order to get work done.⏎Behaviour-driven development (BDD) takes the position that you can turn an idea for a requirement into implemented,
-
smoke,<|endoftext|>What is our proof?⏎The science behind Structured Democratic Dialogues⏎There are two realms for the work of Structured Democratic Dialogues (SDD):⏎The supporting science (known
-
create, so the gift of fire symbolises kinship between mortals and gods.⏎However, in Hesiod’s telling of the story (in his Theogeny), the earliest full account of the story of Prometheus, man already had possession
-
The structure of a story⏎BDD provides a structure for a story. This is not mandatory – you can use a different story format and still be doing BDD – but I am presenting it here because it has been proven to work on many
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
- L1:431 suppresses
- L0:1605 weakly suppresses
- L2:2176 weakly excites
- L1:828 weakly excites
- L2:1184 weakly excites
- L2:92 weakly excites
- L1:1303 weakly excites
- L2:2482 weakly suppresses
- L1:2350 weakly excites 'reasons' or 'permitting' after a qualifier
- L2:1275 weakly excites
- L1:2441 weakly suppresses
- L0:2683 weakly suppresses
- L2:2993 weakly suppresses
- L2:3026 weakly suppresses
- L1:927 weakly excites
- L0:1273 weakly excites
- L1:2407 weakly excites
- L2:2944 weakly excites
- L2:2735 weakly excites
- L0:324 weakly excites
- L2:1789 weakly suppresses
- L2:2465 weakly suppresses
- L2:2236 weakly suppresses
- L0:853 weakly excites
Feeds later layers
- L7:1963 weakly excites
- L9:441 weakly suppresses
- L11:1113 weakly excites
- L15:1759 weakly suppresses
- L4:1529 weakly excites
- L11:2765 weakly excites
- L12:2793 weakly suppresses
- L25:1277 weakly suppresses
- L9:1806 weakly excites
- L13:1633 weakly suppresses
- L7:2920 weakly excites
- L4:817 weakly suppresses
- L13:2644 weakly excites
- L14:233 weakly suppresses
- L10:583 weakly excites
- L6:810 weakly excites
- L4:44 weakly excites
- L20:1785 weakly excites
- L20:2365 weakly excites
- L16:1706 weakly suppresses
- L11:1619 weakly suppresses
- L24:133 weakly suppresses
- L12:2929 weakly suppresses
- L23:1713 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
- {"
- 长
- ante
- physics
- 瘤
- chemistry
- 物理
- nte
- 長
- 迟到
Suppresses
- ��
- urnished
- Clickable
- veillance
- actly
- TextStyle
- agini
- pressions
- conditionally
- ollipop