unclear: word pieces inside names
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 fairly consistent: several of them 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 fires on a piece of a word or name, such as the 'State' in 'Penn State' or the 'acin' in 'niacin'. It is hard to find a single theme, and turning it up did not produce a clear change in what the model wrote.
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.
-
Add your thoughts below. You can sign in with Google, Twitter, Facebook, Yahoo! to get your story posted straight away - or just post a comment to be published later.⏎Let us help you plan
-
health is that severe deficiency of certain vitamins can lead to well-known diseases such as:⏎- Rickets caused by vitamin D deficiency⏎- Pellagra caused by niacin deficiency⏎- Scurvy caused by vitamin C deficiency⏎- Beriberi
-
drive to the nearest health clinic to visit a doctor. In rural South Africa, it is often times more practical for the care to come to the patient. Penn State engineering students work in South Africa to bring improvements to these mobile medical clinics.⏎This
-
14, mostly due to the fact that the United States and the Ottoman Empire had chosen opposite fronts in the doom of World War I.⏎A Turk named Yunus Yaşo, who did not want to be buried in a Christian cemetery, opened
-
the forests of Pennsylvania as “the natural produce of the country.” Consider using native trees in yards to preserve our habitats.⏎For additional assistance, contact your local Penn State Master Gardeners or local Native Plant Society.⏎Return to Pest Pages⏎Gypsy
-
Egg-hatching season is here, and that has some people on edge,” said Emelie Swackhamer, a horticulture educator with Penn State Extension. She added that the pest now has been reported in 3
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 3D modeler and a 3D modeler who has created a model to simulate the motion of the world. I have to get a 3D modeler to simulate the motion of the world. I have
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
by the Young Men's Republican Union.⏎To some it seemed a strange invitation; Peter Cooper had served as a Jacksonian Democrat on the New York Board of Alderman. But he was also an ardent enemy of slavery and he was glad to
-
, he made sounds similar to that of a horse (ashwa). Hence he was named as Ashwathama. Drona always had a soft corner for his son. Dron
-
una in archery, Bhima and Duryodhana in fighting with mace and so on.⏎Dronacharya had a son by name Ashwathama. When Ashwathama was born
-
invoked His divine weapon, the Sudarshana Chakra. It immediately appeared in front of Him in all its brilliant lustre. Krishna then asked Ashwathama to lift the Chakra and take with him if he could.⏎Ashw
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
- L0:178 suppresses
- L0:1224 suppresses
- L0:1301 suppresses
- L0:1 suppresses
- L0:1329 weakly suppresses
- L0:201 weakly excites
- L0:2146 weakly excites
- L0:868 weakly excites
- L0:2201 weakly suppresses
- L0:1670 weakly excites
- L0:77 weakly excites
- L0:2995 weakly suppresses
- L0:1622 weakly suppresses
- L0:319 weakly excites blank space after a label colon
- L0:2259 weakly suppresses
- L0:1565 weakly suppresses
- L0:571 weakly excites
- L0:1075 weakly excites
- L0:1161 weakly excites
- L0:2599 weakly suppresses
- L0:799 weakly suppresses
- L0:1461 weakly suppresses
- L0:2189 weakly excites
- L0:1497 weakly excites
Feeds later layers
- L22:1463 weakly suppresses
- L5:2145 weakly suppresses
- L23:2451 weakly suppresses
- L17:1594 weakly excites
- L17:491 weakly excites
- L6:876 weakly excites
- L6:1671 weakly excites
- L24:1016 weakly suppresses
- L23:2829 weakly excites
- L13:931 weakly suppresses
- L17:627 weakly excites
- L22:505 weakly excites
- L2:398 weakly suppresses
- L3:2634 weakly excites
- L23:299 weakly suppresses
- L22:1369 weakly excites
- L10:411 weakly excites
- L23:239 weakly suppresses
- L6:257 weakly excites
- L9:1404 weakly excites
- L8:605 weakly excites
- L23:1012 weakly suppresses
- L12:1354 weakly excites
- L3:2897 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
- umber
- ثن
- ZN
- 批
- SY
- glam
- ICH
- ST
- reck
- 特殊情况
Suppresses
- IVITY
- odia
- peria
- pagina
- onas
- 的压力
- PRESSION
- imulation
- haar
- pressed