'unless' after a prohibition
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. 6 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 responds to the word 'unless' when it comes right after a rule or prohibition, as in 'should not be used unless' or 'do not adjust your dose unless'. Its strongest negative responses are on parts of Chinese names like 'China' and 'Chiang'. Turning it up did not bring out a clear theme in the sample text.
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.
-
manager from this point to ensure the employee’s contributions are valued and that their needs in knowing how to do their job are<|endoftext|>No person shall hold any civil office unless that person be a qualified elector for such office. Article IX, Section
-
women who weigh more or who exercise regularly.⏎Very low calorie diets of fewer than 800 calories per day should not be used unless you are being monitored by your doctor.⏎Healthy Weight Tip⏎Set goals to eat
-
medical condition being treated⏎- Other medical conditions you may have⏎- Other medications you may currently be taking.⏎As with any medication, do not adjust your dose unless your healthcare provider specifically instructs you to do so.⏎(Click Monopril
-
recommended, this is the final policy that was adopted by our pack.⏎Pack Technology Policy⏎Cub Scouts should not use electronic devices at Cub Scout meetings and events unless the activity requires the use of them
-
the same principle as the coil in an electric heater. There is a black spot on the metal at the back of the stove. Water heater elements must never operate unless immersed in water. A sudden voltage increase to the element caused by a power surge
-
resident for the four years immediately preceding election or appointment. One must also be at least twenty-five years old upon assuming office. No person can serve as attorney general unless he or she has been an active member of the State Bar of Georgia for at
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: the question is to find the answer in a certain way. The answer is to write the code as a program that reads the input, and then outputs the function. The first argument is the list of numbers, and the second argument is the
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
years, Chinese currency existed in both Ancient and Imperial China. In 1914, the Silver Dollar was established as the official currency of the Republic of China, with copper, fen, and nickel coins being added in the 19
-
ung 41349, Taiwan (corresponding author's e-mail: firstname.lastname@example.org, email@example.com).⏎Chia-Liang Chiang is with
-
e-mail: firstname.lastname@example.org).⏎Cite:Ming-Ju Lee, Ming-Gin Lee, Yishuo Huang, and Chia-Liang Chiang, "Purification Study of Pervious Concrete Pavbement," International
-
use synchrotrons – extremely powerful x-ray sources – to trace the complex and intricate structure of the brain.⏎Professor Yeu-Kuang Hwu from the Institute of Physics, Academia Sinica and Professor Ann-Shyn Ch
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
- L4:2588 excites
- L3:1396 excites
- L5:64 weakly excites
- L2:1488 weakly excites
- L5:2650 weakly excites
- L5:2497 weakly suppresses
- L5:2343 weakly suppresses the word "hair"
- L3:2315 weakly suppresses
- L5:102 weakly excites
- L1:2983 weakly excites
- L3:2973 weakly excites
- L4:1209 weakly suppresses
- L5:1462 weakly excites
- L4:577 weakly suppresses
- L2:126 weakly excites
- L4:2852 weakly excites
- L2:1766 weakly excites
- L4:3014 weakly suppresses
- L5:1023 weakly suppresses
- L4:1932 weakly suppresses
- L2:2700 weakly suppresses
- L0:3006 weakly excites
- L3:14 weakly excites
- L2:2763 weakly suppresses
Feeds later layers
- L8:2581 suppresses
- L9:1462 suppresses
- L16:219 weakly suppresses
- L7:829 weakly excites
- L14:223 weakly suppresses
- L8:1147 weakly excites
- L12:2987 weakly excites
- L11:1184 weakly suppresses
- L26:691 weakly excites
- L8:2418 weakly excites
- L7:2145 weakly suppresses
- L16:2891 weakly suppresses
- L19:243 weakly suppresses
- L23:2650 weakly suppresses
- L24:65 weakly suppresses
- L12:249 weakly suppresses
- L20:111 weakly suppresses
- L10:879 weakly suppresses
- L23:1146 weakly excites
- L9:1331 weakly suppresses
- L13:135 weakly suppresses document boundary end-of-text token
- L14:2771 weakly excites
- L7:444 weakly suppresses
- L13:1448 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
- 艿
- esson
- inston
- byname
- &W
- mascul
- USE
- 对其
- teams
- helmet
Suppresses
- 代
- idend
- iversal
- ####⏎
- fers
- fections
- ected
- ],[
- itud
- cles