the word 'gene' and relatives
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. 4 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 'gene' and words built from it, such as 'genealogy' and 'genetic'. It is a clear topic detector, but turning it up or down did not change the text in an obvious, meaningful way.
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.
-
ius' genealogy began in 1996 and is the fifth of its kind.
-
about how your disease will progress and more information about the risk of eye disease for other members of your family. Knowing the gene that causes your eye disease may also be required to be eligible for some types of
-
“pure” version of the diet.⏎Few published studies evaluate risks and benefits of a raw food<|endoftext|>Females descendants of Confucius will be included in the genealogy now being compiled.⏎The current project to trace Confuc
-
6, corn containing a gene that allows it to create a protein that is toxic to certain insects, yet safe
-
get a mutation test done,” said Axelrod. “This is important knowledge for his family to have, especially if he has kids.” She said that a positive gene mutation test also increases the chances of getting prostate and pancreatic cancer.⏎A family history
-
a gene being made. Those duplicates can be passed along to subsequent generations and, on very
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 a 3D object, and I have a 2D object. What is the result of this? The answer to the question is that, and the object has a 3D object. I need to make sure that
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
bodies of water, sometimes in small groups called a waterspout cluster or family. Unlike tornadoes over land that often become visible due to the dust and debris present within them, a waterspout is made up of swirling droplets of water
-
determining, and to appreciate the ever-present urgency to engage these problems and develop ethical and effective ways to restore and sustain the
-
in other words they are actually involved in the dispute that appears before them; or having a pecuniary or proprietary interest in the outcome of the decision, which means they stand to lose or gain (power, influence, money or other
-
of the Central Pacific and the governor of California, deposited the first shovelful of dirt for the railroad’s embankment …”⏎Designs were in place for the Sacramento shops by 1867 and two years later, by
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
- L3:159 excites
- L3:615 excites
- L3:249 suppresses
- L3:805 weakly excites
- L1:2708 weakly suppresses
- L3:1408 weakly excites
- L2:1862 weakly excites
- L3:1176 weakly suppresses
- L1:1198 weakly suppresses
- L3:113 weakly suppresses
- L3:691 weakly excites
- L0:1604 weakly suppresses
- L2:2789 weakly excites
- L3:995 weakly suppresses
- L1:1630 weakly excites
- L1:1971 weakly suppresses
- L1:431 weakly suppresses
- L2:2934 weakly excites
- L3:2197 weakly excites
- L2:2649 weakly suppresses
- L2:1026 weakly suppresses
- L1:169 weakly suppresses
- L3:1141 weakly suppresses
- L1:1670 weakly excites
Feeds later layers
- L8:423 excites
- L5:667 excites
- L25:2750 suppresses
- L11:173 excites
- L5:10 excites
- L8:852 excites comma inside large numbers
- L16:592 weakly excites
- L18:708 weakly excites
- L16:414 weakly excites
- L25:1030 weakly excites
- L17:259 weakly excites
- L23:883 weakly suppresses
- L21:1905 weakly suppresses
- L7:116 weakly excites
- L18:309 weakly suppresses
- L20:2979 weakly suppresses
- L17:2009 weakly suppresses
- L24:1998 weakly suppresses
- L9:2231 weakly excites
- L22:1515 weakly suppresses
- L26:2939 weakly excites
- L11:611 weakly suppresses
- L22:1970 weakly suppresses
- L12:2575 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
- akter
- Investigators
- estation
- 킷
- erculosis
- Preference
- ulatory
- fare
- iscing
- ница
Suppresses
- ése
- öt
- dead
- 对他们
- grav
- +Sans
- sizes
- 擒
- xpath
- deduct