just before a document ends
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. 21 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 last word of a web page or article, right before the text break that marks the end of one document and the start of another. Turning it up produced rambling, repetitive text rather than anything about endings.
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.
-
220.127.116.11, you can define a List of Users and Cities. Create a new variable and:⏎Just<|endoftext|>George Washington was faced with many challenges during his presidency as he not only had to
-
State, and Cornie Scheffer, professor of biomedical engineering at Stellenbosch University, served as mentors on this project.⏎A report on Penn State Live<|endoftext|>You might have seen a vaping advert, whether you are a parent with a child
-
Chile to speed up its judicial process and create a national plan to search for disappeared persons.⏎From Jurist, Aug. 23. Used with permission.<|endoftext|>Oise River, river that rises in Belgium in the Ardennes mountains, southeast
-
These snakes are harmless to humans and are helpful to keep rodent populations in check. If you see one, just let them be and wave from a distance.<|endoftext|>How to Read Diode Numbers⏎By James Miller⏎The system for categorizing
-
(picking which cells are initially turned on) and watching the world behave accordingly to the set of rules.⏎View/Play Conway's Game of Life now!<|endoftext|>It may still be difficult for some people to accept that an activity such as gambling
-
9", "crochet school", "learn to crochet", "treble crochet", crochet⏎| Digg This⏎| Save to del.icio.us<|endoftext|>Back in the early 20th century, the mountains and hollers
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: how to make a subject to the right size to the light in the dark to the environment to the health to the environment to the health to the environment to the health to the environment to the health to the environment to the health to the environment
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
only consumes constant space.⏎In an imperative language (here Modula-3) the same would be achieved with the following code:⏎TYPE List = REF RECORD next: List; value: INTEGER; END;⏎VAR x := NEW(List, value:=
-
and Physiology in Health and Illness 12e⏎Goodreads helps you keep track of books you want to read. Want to Read saving…. Want to Read Currently Reading Read. Other editions. Enlarge cover. Error rating book.⏎Browse
-
C and Immune Function, 2017, https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5707683/#:~:text=It%20is%20a%20
-
ATRIP complex induces the Fanconi Anemia repair and Nucleotide Excision Repair (NER) pathways, as well as the p53 signaling pathway. C. Finally, a
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
- L5:601 strongly suppresses
- L6:449 strongly excites
- L16:37 strongly excites
- L18:555 strongly suppresses end-of-text marker between documents
- L18:221 suppresses
- L0:54 weakly suppresses
- L19:197 weakly suppresses
- L19:798 weakly suppresses
- L4:2339 weakly suppresses
- L20:299 weakly suppresses
- L19:283 weakly suppresses
- L3:50 weakly excites
- L20:1398 weakly suppresses
- L4:97 weakly suppresses
- L20:249 weakly suppresses
- L20:221 weakly suppresses
- L10:3 weakly excites
- L18:931 weakly excites
- L20:2371 weakly excites
- L5:2458 weakly excites
- L4:804 weakly suppresses
- L16:454 weakly suppresses
- L1:141 weakly excites
- L16:93 weakly suppresses
Feeds later layers
- L25:436 suppresses
- L27:1300 suppresses
- L22:230 suppresses
- L22:133 weakly excites
- L26:1349 weakly suppresses
- L27:158 weakly excites
- L23:2081 weakly suppresses
- L27:10 weakly excites
- L26:2922 weakly suppresses
- L26:1356 weakly excites
- L27:1962 weakly suppresses
- L26:2341 weakly excites
- L23:2483 weakly excites
- L24:264 weakly excites
- L27:2050 weakly suppresses
- L27:2818 weakly suppresses
- L25:2777 weakly suppresses
- L25:939 weakly suppresses
- L23:2142 weakly excites
- L22:2283 weakly suppresses
- L25:827 weakly suppresses
- L22:2324 weakly suppresses
- L25:2135 weakly excites
- L22:300 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
- UNCH
- zap
- van
- amb
- d
- files
- endet
- ucumber
- -files
- van
Suppresses
- 干什么
- ares
- iais
- theless
- asily
- 的主题
- 勾
- uble
- 徂
- semiclass