percent-encoded url characters
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. 1 of Qwen3-0.6B. Its strongest snippets are mixed: they peak on different tokens. 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 reacts to the leftover characters of web addresses and encoded text, such as the '0' in '%20' or the 'F' in '%2F'. Turning it up made the model write about unrelated things, so the steering gives no clear picture.
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.
-
017, https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5707683/#:~:text=It%20is%20a%20potent%20antioxidant
-
and%20adaptive%20immune%20system.⏎- How Vitamin C Supports a
-
7075. . . http://www.plosone.org/article/info%3Adoi%2F10.1371%2Fjournal.pone.0057075.<|endoftext|>Racism
-
archive.jpr.org.uk/10.1177%2F1359104505053754
-
- Understanding a patient's values and behaviors is seen as one way to reduce health disparities⏎- Cultural competency acknowledges that provider and patient may have different goals%2C experiences⏎- Race and Health Disparities workgroup of FR%3DE
-
) is a proto-oncogene that encodes a protein known as hepatocyte growth factor receptor (HGFR). The hepatocyte growth factor receptor protein possesses tyrosine-kinase activity
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 had a lot of time to think about it" and that it was something I had to think about it. So when I think about it, I had to think about it. But it was something I had to think about it
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
from a blog posting entitled Disk Partition Alignment Best Practices for SQL Server by Jimmy May and Denny Lee at Microsoft.⏎You have a server with a RAID volume – For the example we will say RAID 1
-
6 bytes, exactly 1024 KB or 1 MB. This is a much better solution because it scales well across the most common stripe sizes ââ‚<|endoftext|>Many developing countries rely on traditional medicine as an accessible and affordable treatment
-
(core) city ; Greek ἄστυ asty ) with the associated surrounding area ( χώρ
-
ν ἄνθρωπον ἐγένετο καὶ οὐχ ὁ ἄνθρωπος διὰ τὸ σάββ
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
- L0:1301 suppresses
- L0:1224 suppresses
- L0:178 suppresses
- L0:1 weakly suppresses
- L0:1849 weakly suppresses
- L0:2809 weakly excites
- L0:961 weakly suppresses
- L0:2082 weakly suppresses
- L0:19 weakly excites
- L0:29 weakly suppresses
- L0:1229 weakly suppresses
- L0:1591 weakly suppresses
- L0:2628 weakly suppresses
- L0:2907 weakly excites
- L0:652 weakly suppresses
- L0:365 weakly excites
- L0:1021 weakly excites
- L0:2433 weakly suppresses
- L0:202 weakly excites start of a new document
- L0:2299 weakly suppresses
- L0:237 weakly suppresses
- L0:61 weakly excites
- L0:920 weakly suppresses digit '9' inside years and numbers
- L0:2422 weakly excites
Feeds later layers
- L9:2150 excites
- L22:550 weakly excites
- L12:2858 weakly excites comma or bracket after a name in a byline
- L15:1415 weakly excites
- L9:13 weakly excites
- L8:2221 weakly excites
- L9:2572 weakly excites
- L20:2851 weakly excites
- L23:1050 weakly excites
- L2:2688 weakly suppresses
- L22:620 weakly suppresses the word 'if'
- L18:976 weakly suppresses
- L19:1893 weakly suppresses
- L23:575 weakly excites
- L4:1064 weakly suppresses
- L22:488 weakly excites
- L9:2616 weakly excites
- L13:2005 weakly excites
- L9:2258 weakly suppresses
- L16:1944 weakly suppresses
- L12:2810 weakly excites
- L22:1577 weakly excites
- L25:349 weakly excites
- L11:2721 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
- ansom
- ÃO
- <const
- Resolved
- ��
- Requires
- balls
- ancia
- ônica
- oner
Suppresses
- 事
- ạt
- center
- hid
- 事迹
- 事后
- Mouth
- /Internal
- mouth
- -cache