'because' after a claim
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. 0 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 the word 'because' when it comes right after a statement and introduces the reason for it. Turning it up made the model's text drift into general, unrelated wording, so the steering did not reveal 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.
-
ists. They use modern conservation practices to protect our nation's water supplies. Many times these efforts are put in place voluntarily because farmers are driven by a strong stewardship ethic.⏎However, the restoration bill largely
-
month chosen because of a widely believed tradition that the world was created in March, and November is
-
about the fate of your tree, carefully look at the extent of the damage and assess the overall health of the tree.⏎“It could be that the tree was damaged because it’s diseased, weak or on its way out anyway,” Trinklein
-
-rap had an intellectual bent because of accessibility of education which was promoted by policies like the affirmative action, it’s
-
Mechanical strength is critical because the shell protects the instruments inside against mechanical shock. Enclosures for on-whale
-
are vital because they help train kids for a new frontier of jobs, Insight Communications Director of Communications
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 reading this for a long time, I have been thinking about it. I have been thinking about it. I have been thinking about it. I have been thinking about it. I have been thinking about it. I have been
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
ottled if you turn the rotating stage|⏎|Tourmal<|endoftext|>The Catechism of the Catholic Church states:⏎"Every act of war directed to the indiscriminate destruction of whole cities or vast areas with their inhabitants is a crime against God
-
of urgency for a patient’s ocular findings such as visual field defects, cranial neuropathies, double vision, optic neuropathy, ptosis, pupillary abnormalities and loss of vision.⏎How long does it take for vision therapy to work
-
”在汉语中具有较大的作用,使用率高,但是泰国学生经常在学习和使用汉语助动词中产生各种的偏误。本文通过调查法、统计
-
't have family or friends to turn to.⏎Barack Obama the president of the United States has a plan.'⏎The people described above are America's homeless.These unfortunate individuals often we see living on the streets in the downtown areas of our
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
None: this is the first layer.
Feeds later layers
- L10:851 weakly excites
- L9:1002 weakly suppresses comma after who, when or where
- L6:313 weakly suppresses
- L23:706 weakly suppresses
- L18:2010 weakly suppresses
- L22:258 weakly suppresses
- L21:1336 weakly suppresses
- L10:571 weakly excites
- L25:1205 weakly suppresses
- L4:3039 weakly suppresses
- L7:2934 weakly suppresses
- L5:2009 weakly suppresses
- L7:1877 weakly suppresses
- L18:2890 weakly excites
- L18:1785 weakly suppresses
- L3:640 weakly excites
- L6:665 weakly excites
- L21:1605 weakly suppresses
- L6:178 weakly excites
- L22:627 weakly suppresses
- L26:2935 weakly suppresses
- L21:821 weakly excites
- L5:940 weakly excites
- L7:379 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
- nes
- nowhere
- tering
- ippo
- Associates
- luk
- NES
- never
- lee
- PROP
Suppresses
- ;br
- 矾
- щит
- 雨水
- izr
- 节
- 即可
- 时节
- 我可以
- elize