document end token
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. 13 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 fires on the marker that separates one unrelated document from the next, especially after short lists or fragments of page text. It switches off on words cut in the middle of a phrase. Steering did not produce a clear theme.
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.
-
velocity (km/s)...?⏎- Inversely proportional⏎- Directly proportional⏎- Linearly related⏎- Non-linearly related⏎Similar Physics resources:<|endoftext|>RNA functions as the central conduit for information transfer in biology - in simple replicating
-
n");⏎- Compare two blocks of memory (function⏎- Locate first occurrence of character in string (function⏎- Locate last occurrence of character in string (function<|endoftext|>- Mattress Accessories⏎- Sleep Products⏎Also referred to as “responsible
-
, that's possible. 524 = 512.⏎The answer is 214 + 314 = 512<|endoftext|>[Haskell-beginners] Question about data structures⏎daniel.is.fischer at
-
, but pointers can be uninitialized, which is a nasty security hole. An eagerly evaluating functional language without hacks, would only allow for acyclic graphs of pointers.<|endoftext|>The Death And Life Of The Great Lakes Pdf⏎- and pdf⏎- Tuesday
-
⏎July 08, 2009 |<|endoftext|>Computer Science Concentration for Students Matriculated After Jan. 1,
-
Mentor: Carl Ruetz⏎Page last modified July 14, 2009<|endoftext|>what does hamlet think about suicide? with textual support⏎2 Answers | Add
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: what can I do to do it, can I do it. What can I do? What can I do? Maybe that's a problem. Let me try to think of different things. What can I do, maybe I can do that
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
old stick to bash natural gas with is that nations will burn coal instead. Or just ignore the zealots and move forward rapidly to produce, export, and import methane. While Canada’s elite stands there slack-jawed going but what about our
-
child dies of a preventable cause, it is a huge tragedy and the nation has to get its act together and intensify efforts to make sure it doesn’t happen,” Anuradha Gupta, a senior official at the Health Ministry, said
-
percent of youths have witnessed cyberbullying and 49 percent of students have regretted something they have posted online.2 Girls more often resort to cyberbullying while guys still rule the physical bullyingworld on playgrounds and in school hallways
-
sorted and if not swapping randomly two elements. This is repeated until eventually the sequence is sorted.⏎If we consider all permutations of the first 4 natural numbers as input the expectation value of the number of swaps, averaged over all 4! input
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
- L12:60 suppresses
- L8:1198 suppresses
- L12:262 weakly excites
- L12:1085 weakly excites
- L10:63 weakly excites
- L11:2128 weakly suppresses
- L11:1672 weakly excites
- L12:328 weakly suppresses
- L12:800 weakly excites
- L10:408 weakly excites
- L7:716 weakly excites
- L12:2052 weakly excites
- L9:2023 weakly suppresses
- L12:594 weakly excites
- L12:151 weakly suppresses
- L8:888 weakly excites
- L12:254 weakly suppresses
- L11:526 weakly excites
- L2:310 weakly suppresses
- L9:2275 weakly excites
- L11:1414 weakly excites
- L6:306 weakly excites
- L0:2805 weakly suppresses
- L9:40 weakly suppresses
Feeds later layers
- L14:1917 suppresses
- L14:179 excites
- L14:1660 suppresses
- L14:806 suppresses
- L14:1543 suppresses
- L14:256 suppresses
- L14:3011 suppresses
- L15:2488 excites
- L14:2133 suppresses
- L16:1425 suppresses
- L17:2879 suppresses
- L14:697 excites
- L14:2331 excites
- L14:1745 excites
- L15:2519 suppresses
- L15:2468 weakly excites
- L14:358 weakly suppresses
- L14:1902 weakly excites
- L15:2479 weakly suppresses
- L17:2499 weakly suppresses
- L14:2241 weakly suppresses
- L16:2497 weakly excites
- L14:2947 weakly suppresses
- L15:1147 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
- ##⏎⏎
- ',⏎⏎
- życie
- ’⏎⏎
- validated
- $⏎⏎
- ’.⏎⏎
- ';⏎⏎
- ';⏎⏎
- itle
Suppresses
- _periods
- 这方面
- 前三季度
- addObject
- ewire
- .getRight
- _REUSE
- enschaft
- 左右
- createForm