end of document, and 'it turned out'
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. 19 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 mostly reacts to the marker that signals the end of one document, and it also reacts weakly to the phrase 'it turned out'. Its steered outputs were unrelated and broken, so the effect of turning it up or down is hard to read.
Written by Claude from this neuron's activations and steering outputs. Confidence: low. 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.
-
.⏎ASJC Scopus subject areas⏎- Electronic, Optical and Magnetic Materials⏎- Ceramics and Composites⏎- Condensed Matter Physics⏎- Materials Chemistry<|endoftext|>Humans are intimately connected with the physical environment, even though we have only been present
-
if you need to, you can always buy college assignments online.<|endoftext|>Breaking research news from a team of scientists led by Carnegie's Ho-kwang "
-
grammatical functions and features of Chinese auxiliary verbs.<|endoftext|>Obviously, the unlimited liability feature of general partnerships is a serious impediment to conducting
-
and ultimately foster innovation.<|endoftext|>Resource Center: Allergy Shots (Immunotherapy)⏎For people suffering severe allergic
-
Issues⏎Being able to troubleshoot common HVAC issues can be beneficial<|endoftext|>Whether intentional or not, if this process continues it will complete, perhaps finally,
-
support your existing curriculum, depending on your needs. There are cross-curricular links and activities suggested throughout.<|endoftext|>Usually, consumers don’t have problems with online financial transactions. However, occasionally something
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: the number of, I know, the number of, the number of, the number of, the number of, the number of, the number of, the number of, the number of, the number of, the number of, the
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
produce structures that are more efficient to run, cutting ongoing costs and minimizing the carbon footprint.⏎What does this have to do with BrightBuilt Barn?⏎Well, it turns out that the Barn was also designed with (simpler and much less expensive
-
to know how EGT1 helps roots to overcome gravity. For this, the researchers looked at the genes disrupted in the EGT1 mutant. This, it turned out, were genes for producing parts of the cell wall. This disruption translated into
-
the work of a small team at IBM in Switzerland that had developed a complex technique to image or “take a picture” of molecules in the journal Nature — it turned out to be the coup de grace for the team.⏎Back in 20
-
searched for a mutant in barley that didn’t do this. And successfully so. They found a plant whose roots grew more downwards than sideways. This mutant, it turned out, had a mutation in the gene EGT1.⏎EGT1 organ
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
- L18:824 suppresses
- L18:30 excites
- L17:2988 weakly excites
- L5:2896 weakly suppresses
- L2:94 weakly suppresses
- L18:429 weakly suppresses
- L15:376 weakly suppresses
- L18:20 weakly suppresses
- L6:2960 weakly suppresses
- L0:2926 weakly excites
- L2:2016 weakly suppresses
- L18:569 weakly excites
- L15:2776 weakly excites
- L1:2683 weakly excites
- L17:271 weakly excites
- L13:197 weakly excites
- L2:2621 weakly excites
- L8:59 weakly suppresses
- L9:837 weakly excites
- L10:65 weakly excites
- L7:1335 weakly suppresses
- L0:2329 weakly excites
- L4:2620 weakly excites
- L1:1464 weakly excites
Feeds later layers
- L23:236 strongly excites
- L26:122 strongly excites
- L20:841 strongly excites
- L27:1266 excites
- L25:4 suppresses
- L24:1896 excites
- L25:745 excites
- L25:430 excites
- L20:1313 excites
- L27:574 excites 'Aut' starts and colon prediction
- L24:1466 suppresses
- L26:1229 excites
- L26:1329 excites
- L27:646 excites
- L27:1695 excites
- L25:440 excites
- L20:1077 excites
- L27:858 weakly excites
- L26:648 weakly suppresses
- L22:940 weakly suppresses
- L24:563 weakly suppresses
- L23:275 weakly suppresses
- L26:485 weakly excites
- L23:587 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
- ":
- :
- ':
- ”:
- ###
- ():
- ):
- ’:
- **:
- \":
Suppresses
- andel
- 不了
- ?"
- inho
- inne
- #$
- #"
- rones
- вис
- atel