preposition after a state word
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. 24 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 a small preposition such as 'in', 'at' or 'of' when it follows a word describing a state, as in 'stay in darkness', 'are at rest' or 'in good health'. Turning it up made the model repeat odd words like 'hushed' and 'hibachi', so the steering result is not clearly related.
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.
-
affected is in our brain - our circadian rhythms.⏎Although the urge to hibernate during the winter months may be strong, humans aren't really meant to stay in darkness for months at a time. We need sunlight - if we get too little
-
. Currently, it has been checked up to the power 4.6 ⋅<|endoftext|>Picture Yourself in Orbit: Are You Really Weightless?⏎An object in orbit is constantly falling, and falling is what causes “weightlessness.” Gravity acts
-
to the imbalances in the mind, spirit, and the body. If these elements are in harmony, you will remain in good health. However, if there is something that
-
nears the end of one of ...Read More⏎Uneven rainfall across the state has helped replenish Northern California reservoirs, while those to the south remain in mostly abysmal shape. Meantime, the latest snow report is in
-
was twice as frequent in those who had intermediate cardiovascular health and two and a half times more frequent in those who had poor cardiovascular health when compared to those who were in good health. With anxiety disorder, the frequency of depression was 20%
-
destroy tissue and release energy, thereby generating heat. Our metabolism speeds up when we exert ourselves, and we burn more calories. Our metabolism slows down when we are at rest, and we burn fewer calories. But even at rest we continue to expend
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: [t]d endt endt endt endt endt endt endt endt endt endt endt endt endt endt endt endt endt endt endt endt endt endt
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
It may cause you to feel sleepy, especially when you first start taking it.⏎|Type of medicine||An antiepileptic medicine|⏎|Available as||Tablets and oral liquid medicine (suspension)|⏎Having epilepsy means that you
-
) is a proto-oncogene that encodes a protein known as hepatocyte growth factor receptor (HG
-
ators. The less time patients spend in there the better off they are.”⏎For patients who spend several weeks in the ICU, that race is sometimes lost to late-onset sepsis, a condition where patients succumb to a systemic infection that
-
not.⏎Taken from…. The Future of Education: Tablets vs. Textbooks http://mashable.com/2012/10/05/tablets-vs-textbooks/⏎They also say…. ‘ Students with special learning needs may
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
- L21:1577 excites
- L7:2333 weakly excites
- L14:2929 weakly excites
- L15:686 weakly suppresses
- L1:44 weakly excites
- L9:2606 weakly suppresses
- L1:624 weakly suppresses
- L15:1842 weakly excites
- L23:1294 weakly excites
- L21:1511 weakly suppresses
- L22:2908 weakly suppresses
- L2:1109 weakly suppresses
- L7:1288 weakly excites
- L23:1343 weakly suppresses
- L22:2377 weakly excites
- L8:2095 weakly suppresses
- L17:278 weakly excites 'as' in 'as well as' and 'as much as'
- L8:1846 weakly suppresses
- L2:570 weakly suppresses
- L18:2425 weakly excites
- L11:1906 weakly excites
- L0:1217 weakly excites
- L15:843 weakly excites
- L15:1951 weakly excites
Feeds later layers
- L27:2112 weakly excites
- L26:1833 weakly excites
- L27:2899 weakly excites
- L25:1591 weakly excites
- L27:654 weakly suppresses
- L25:2473 weakly excites
- L25:976 weakly excites
- L26:977 weakly suppresses
- L26:1881 weakly excites
- L25:2654 weakly excites
- L27:2441 weakly excites
- L25:1262 weakly excites
- L27:2098 weakly excites
- L25:1924 weakly excites
- L25:1940 weakly excites
- L27:2614 weakly excites
- L25:1752 weakly excites
- L27:1164 weakly excites
- L26:269 weakly excites
- L25:1881 weakly excites
- L25:1005 weakly excites
- L27:2488 weakly excites
- L27:2916 weakly excites
- L25:1913 weakly excites 'mil' or 'bil' inside names
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
- 失
- rem
- balance
- BAL
- bal
- Balancer
- 平衡
- memo
- balance
- 開放
Suppresses
- tin
- t
- ニング
- olist
- etting
- panion
- inated
- /xhtml
- àn
- atics