'a' after 'tell', 'bring' or 'much of'
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. 5 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
The neuron fires on the word 'a' in phrases like 'tell a history', 'bring a fresh perspective' or 'much of a workout'. Turning it up did not give the writing a clear new subject.
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.
-
determine the movements necessary to pick up a weight in the invented environment. Grab hold, stiffen your arm, and lift. It doesn't sound like much of a workout, but the researchers say it could be useful in therapy for people recovering from
-
49F for pure metal and about 512F for tin-zinc alloys. Tests show that screws plated with the material do not stand much of a chance in hot areas. One test run several years ago by J. Bick
-
collections in Greater Manchester⏎A compass used in a PoW escape, a railway station sundial and part of the world's first commercial computer: objects that tell a history of Manchester and its place in the world.⏎BBC Manchester and museums across Greater
-
ofs is an introduction to modern symbolic logic, covering sentential and predicate logic (with identity). The course is highly interactive and engaging. It brings a fresh perspective to classical material by focusing on developing two crucial logical skills: strategic construction
-
Kitten? How Do I Tell a Kitten’s Age?⏎If your kitten is yet to get its adult teeth,
-
, must speak to a much wider audience: to the educated general public and academic audiences worldwide.⏎Inquiry editors are members of that wider audience, who can tell an author when a concept, process, or term is or is not clear. Often
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 question about problem of a 2D 2D problem of a problem of a 2D problem of a 2D problem of a 2D problem of a 2D problem of a 2D problem
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
Moving Picture Experts Group⏎||Flexible, works with a wide range of editing and compression programs; have support for multiple video or audio streams⏎||MP4 is are usually compressed; compatible with more devices like iPad, iPod, iPhone, Android Phone
-
. R is a programming language for statistics computing. These tools, especially their open source packages, have
-
muscle tissue and white blood cells of autistic children, as well as in the brains of mice with a different mutation linked to autism2,3.⏎“We’ve had this global sense that mitochondrial dysfunction may be particularly relevant to synaptic defects, but I think
-
done with the bill. See HERE in the case of the Red-breasted Parakeet (Psittacula alexandri). On the other hand, with aboreal birds, they scratch with
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
- L4:137 excites
- L4:2589 suppresses
- L3:1456 suppresses
- L4:52 excites
- L4:731 excites
- L3:2603 suppresses
- L4:2830 excites
- L4:2242 excites
- L4:283 weakly suppresses
- L3:597 weakly suppresses
- L0:1343 weakly excites
- L3:811 weakly excites
- L4:49 weakly excites
- L3:2843 weakly suppresses
- L0:173 weakly excites
- L0:1717 weakly suppresses
- L1:1684 weakly suppresses
- L2:699 weakly excites
- L3:1920 weakly excites
- L2:670 weakly excites
- L0:510 weakly excites
- L2:1446 weakly excites
- L0:2713 weakly suppresses
- L1:1078 weakly suppresses
Feeds later layers
- L16:2434 excites
- L8:1673 suppresses
- L23:728 suppresses
- L21:592 suppresses
- L17:188 suppresses 'a' before a noun phrase
- L6:530 excites
- L16:815 suppresses 'the' before a noun mid-sentence
- L8:466 suppresses 'a' or 'an' before a noun phrase
- L9:59 weakly suppresses "a" meaning per (a day, a year)
- L16:361 weakly suppresses
- L24:923 weakly excites
- L25:390 weakly suppresses
- L26:259 weakly suppresses
- L16:2701 weakly suppresses modifiers right before a noun
- L19:1060 weakly suppresses
- L23:2755 weakly excites
- L16:880 weakly suppresses
- L16:1309 weakly suppresses
- L25:837 weakly suppresses
- L8:595 weakly suppresses
- L26:110 weakly excites
- L20:901 weakly suppresses
- L14:990 weakly suppresses
- L19:1908 weakly suppresses
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
- ingham
- 等活动
- 的脚步
- edImage
- xff
- ingly
- 给出
- 别说
- illions
- 香味
Suppresses
- ignment
- aisal
- ⇑
- sweeping
- reement
- field
- UA
- 弧
- reass
- dynamic