comma before a 'because' or 'as' clause
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. 25 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 a comma that comes right before a reason, such as 'not a substitute, as this will be toxic'. When it was turned up, the model kept writing the word 'as' over and over.
Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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.
-
this happens, add a watering agent such as Water-In to enhance soil penetration. Water-In is available at most garden stores. Do not use detergent as a substitute, as this will be toxic to the plant.⏎Heavy clay soils drain slowly. If
-
organised, and properly targeted for temperature-controlled shipping and storage.⏎First, the purpose of the kit or kits must be firmly defined. This is a critical stage, as imprecision in
-
data, only 2 out of 10 households meet the estimated average requirement (EAR) of Vitamin A in their daily diet. This trend is particularly worrying, given the rising rates of food insecurity that arose as a consequence of the COVID-
-
to wipe the fridge.⏎· Avoid using any soap or detergent, as they might leave behind a residue or scent, which can contaminate your
-
is made up of swirling droplets of water. Many people assume that a waterspout is nothing more than a tornado over water. This is not entirely true, however, as some waterspouts can develop in relatively fair weather conditions unlike the
-
ores⏎This is a fun activity to do at the beginning (or end) of the school year. Don’t attempt this on a cool or overcast day, because you need heat and direct sunlight to cook the s’mores. You will
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: when the pastes and pastes are because of the pastes because of the pastes... or the pastes because of the pastes and the pastes because of the pastes. I'm not sure if this is the right way
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
U.C. Mohanty, a professor in the Centre for Atmospheric Sciences at the Indian Institute of Technology, and Mohanty's doctoral student, Ashish Routray, collaborated with scientists at the National Center for Atmospheric Research and gathered information
-
examples of air pollution.<|endoftext|>Revealing the art and architecture of Vijayanagara Empire, Vittala Temple Complex is one of the major tourist attractions in Hampi. The complex and its structures embody the pinnacle of Dravidian temple architectural
-
D) ministry. Rather, the demand follows from a chest thumping and energetic flag waving ‘Tiranga March’ that was recently held at the Jawaharlal Nehru University (JNU).⏎The
-
. During these ten years he emerged as the overseas leader of the Indonesian nationalist movement and became acquainted with counterparts representing other independence movements, including the Indian leader Jawaharlal Nehru. In 1927 Hatta was accused by Dutch
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
- L22:2261 suppresses
- L22:1385 suppresses
- L22:592 excites
- L21:2115 excites
- L21:1111 suppresses
- L24:951 excites
- L22:2677 excites
- L22:1883 excites
- L22:684 suppresses
- L19:710 suppresses
- L24:1392 suppresses
- L19:1570 excites
- L10:39 excites
- L18:1444 suppresses
- L20:541 suppresses
- L5:136 excites
- L19:2495 excites
- L20:1043 excites
- L21:2669 suppresses
- L23:1635 excites
- L19:1036 suppresses
- L23:986 excites
- L10:190 weakly excites
- L21:2563 weakly suppresses
Feeds later layers
- L26:588 strongly excites
- L27:471 strongly excites
- L26:2423 excites
- L26:790 suppresses
- L27:811 excites
- L26:1024 weakly excites
- L26:2115 weakly suppresses
- L26:2926 weakly suppresses
- L26:1320 weakly suppresses
- L27:1497 weakly suppresses
- L26:2152 weakly suppresses
- L26:220 weakly suppresses
- L27:1676 weakly suppresses
- L27:493 weakly suppresses
- L27:1227 weakly suppresses
- L27:1082 weakly suppresses
- L26:953 weakly suppresses
- L27:1406 weakly suppresses
- L26:2742 weakly suppresses
- L26:2123 weakly suppresses
- L26:218 weakly excites
- L26:2286 weakly excites
- L27:1588 weakly excites
- L27:2961 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
- as
- 作为
- →as
- As
- 作為
- As
- -as
- 作为一个
- as
- _as
Suppresses
- 因为在
- ragments
- because
- 因为你
- ーン
- 因为他
- 因为
- because
- pois
- anchise