'always' or 'often' after 'is' or 'there is'
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. 7 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 words like 'always' and 'often' when they sit between 'is' or 'there is' and the rest of the sentence, as in 'it is always easy'. When it was turned up, the written text did not clearly change in any way tied to this idea.
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.
-
challenging task but with the right anchoring script, it is always easy.⏎So here we will see how to write anchoring script or compering
-
the average rainfall is in upwards of 150″ per year. Temperature varies little, with the average temperature near 82, and the humidity is always high.⏎Figure 2. Yum! After a hard days
-
in the State and nearly 55 persons died because of complications arising out of it. However, sporadic cases continue to get reported from districts and there is always a possibility of the virus making a comeback next
-
to answer the questions that don’t require context first. (You may decide you have to read the entire passage to answer the context questions.)Staying calm is always easier said than done but remember that there is no advantage to having existing subject knowledge
-
, were almost cut off from the rest of Mallorca because of those very mountains. The tortuous, switchback road with steep climbs, hairpin bends and often cut by bad weather conditions made travel so long and arduous to Palma
-
in his time. Muhammad also appears to have been the first to suggest the contagious nature of leprosy, mange and sexually transmitted disease; and that there is always a cause and a cure for every disease, according to several hadiths in
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 to be able to write a 1000-word essay that includes all the content, and I can use any of the content I want, as long as it is in the English language. 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.
-
But these aren’t the kind of rechargeable batteries that you can use in place of disposables.⏎Instead, you might use nickel-cadmium batteries (NiCd). Nickel is the cathode and cadmium is the anode. The recharge
-
Acid Battery: They take longer to lose their charge when not used, making it easier to maintain your battery while away from the camper or RV.⏎Nickel Cadmium Battery: These batteries have been around a long time and can hold a lot of
-
by milling. Whole grains are better sources of fiber and other important nutrients, such as selenium, potassium and magnesium. Whole grains are either single foods, such as<|endoftext|>Of all the
-
blood pressure [*].⏎Focus on foods high in potassium too. Potassium is an electrolyte needed to balance out too much sodium, relax blood vessel walls
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
- L5:599 suppresses
- L3:2798 excites
- L5:2610 excites
- L4:1543 suppresses
- L4:830 weakly excites
- L5:2095 weakly excites
- L0:3009 weakly excites
- L6:1622 weakly excites
- L4:2212 weakly suppresses
- L6:1283 weakly excites
- L5:852 weakly suppresses
- L6:457 weakly excites
- L1:1113 weakly suppresses
- L4:981 weakly suppresses
- L6:2683 weakly suppresses
- L6:1813 weakly excites
- L6:1492 weakly excites
- L3:1386 weakly excites
- L1:2236 weakly suppresses
- L5:1884 weakly suppresses
- L2:1939 weakly excites
- L6:17 weakly suppresses
- L0:37 weakly suppresses
- L4:1787 weakly excites
Feeds later layers
- L8:1056 excites
- L15:1829 excites
- L20:1607 excites
- L17:1603 excites
- L10:932 excites
- L17:1662 weakly excites
- L8:330 weakly excites
- L12:842 weakly excites
- L12:1702 weakly suppresses
- L8:1010 weakly suppresses
- L19:1111 weakly suppresses
- L24:2042 weakly excites
- L12:1320 weakly excites
- L9:1177 weakly excites
- L27:848 weakly suppresses
- L18:1335 weakly suppresses
- L17:1249 weakly excites
- L16:2726 weakly excites
- L20:2576 weakly excites
- L19:1011 weakly suppresses
- L18:2768 weakly suppresses
- L18:1319 weakly excites
- L22:1303 weakly suppresses
- L14:839 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
- TIM
- .Guid
- Editors
- fist
- addElement
- ystone
- ộc
- พฤ
- 交替
- 鹑
Suppresses
- ificate
- het
- ilogue
- htable
- 日正式
- 信息发布
- owe
- 权重
- nowled
- ustomed