word fragments inside longer words
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 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 weakly fires on the middle or end of a longer word that is split up, such as 'ware' in 'Mind Ransomware' or 'istics' in 'Ballistics', and on digits in book ISBN numbers. The activations are small, so it is a weak signal. Turning it up gave unrelated text.
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.
-
not block you from accessing all of your files. However, it can do a lot of other things that can be quite annoying. In fact, Mind Ransomware can basically block you from operating your computer because it really wants you to pay attention
-
size and of a finite energy⏎- Infinitely massive, finitely old, of degenerate energy and uniform⏎1 of 20⏎Other questions in this quiz⏎2. Olbers' Paradox disproved Newton's model by
-
remove this infection from your system. It should be possible to decrypt your files with the key left by the application itself.⏎It is not clear how Mind Ransomware gets distributed around, but since the ransomware is “educational,” it is
-
5⏎Paperback ISBN: 9781503603844⏎Ebook ISBN:
-
2014|⏎|ISBN: ||9781312744967 (DRM-EPUB
-
the comparison of 45-70 vs 444 Marlin, which can be found when clicking that link.⏎45-70 Ballistics Chart:⏎You can use the following ballistics charts as a .45-
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 been so busy writing this as a way of life, but I don't know how to let go of it for myself, and I am so used to the idea that the world is perfect and so I feel that I am in
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
and the Ballotechnic model21 for the limits of complete reaction .⏎- Curious Case of Red Mercury⏎- “The new material is known as a 'ballotechnic' explosive, even though it does
-
that to snowpack,” said Matheus. “Looking at snow, rather than runoff, is an important distinction because it is the source of our water supply.⏎“When we have droughts, they might seem severe on a local time scale,
-
en their back.⏎Dr Lund said: "Scoliosis is a painful illness, and Richard would have been in quite a lot of pain on a daily basis.⏎"These methods could also
-
crossed out with a pen, and the word made department.⏎The rival company acting Torture got away with it successfully, especially, complains our Yorkshire priest "as 'ure' was pronounced 'yaw' in the best southern
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
- L6:120 strongly excites
- L6:599 suppresses
- L3:92 excites
- L6:174 excites
- L2:240 weakly suppresses
- L3:1388 weakly excites
- L0:364 weakly suppresses
- L4:1124 weakly suppresses
- L5:2548 weakly excites
- L0:1067 weakly suppresses
- L5:2685 weakly excites
- L3:2005 weakly suppresses
- L3:103 weakly excites
- L0:2726 weakly suppresses
- L2:642 weakly excites
- L6:362 weakly suppresses
- L5:2725 weakly suppresses
- L1:197 weakly suppresses
- L1:1792 weakly excites
- L4:2091 weakly suppresses
- L0:2209 weakly suppresses
- L4:15 weakly suppresses
- L2:461 weakly suppresses
- L5:2113 weakly suppresses
Feeds later layers
- L8:2615 suppresses
- L19:1801 weakly suppresses
- L16:957 weakly suppresses
- L21:1246 weakly excites
- L9:915 weakly excites
- L23:2878 weakly excites
- L26:103 weakly excites blank space before a number
- L24:2827 weakly suppresses 'as' in 'as far as' and 'as many as'
- L26:1027 weakly excites
- L19:1406 weakly excites
- L13:1149 weakly suppresses
- L14:238 weakly suppresses
- L21:1189 weakly excites
- L27:1453 weakly excites
- L17:449 weakly suppresses
- L26:1268 weakly excites
- L26:2975 weakly excites
- L24:656 weakly excites
- L14:498 weakly suppresses
- L27:2879 weakly suppresses
- L26:2551 weakly suppresses
- L20:1428 weakly excites
- L26:2629 weakly excites
- L27:739 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
- *)⏎⏎
- ertiary
- handleMessage
- 们的
- [e
- mpz
- )["
- [res
- 省教育
- ("?
Suppresses
- ?,?,?,?,
- صل
- leaflet
- kiên
- fabric
- worm
- imension
- réalisé
- 共
- fold