"set" before out, off or up
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. 23 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 word "set" when it is followed by a direction word, as in "set out", "set off" or "set up". Turning it up made the model repeat "off" and "about", which matches the words it helps the model guess next.
Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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.
-
hand. He immediately visited the hospital alongside mechanical engineer Alessandro Romaioli to see the valve themselves, upon learning about the situation. Afterward, the pair set about creating a prototype of the valve to test on the patients. Speaking to BBC
-
9. Can you give Pirate Pete directions to find Percy Parrot?⏎Great work, Shipmates! Now Pirate Pete has gathered his things, he is ready to set sail.⏎Your teachers are sooo proud of you, keep up the good work
-
tied to the idea that farmers who sowed seeds around the full moon had a better harvest.⏎So, over the years, scientists–a naturally curious bunch–set out to see if there is any true link. And what did they find?⏎
-
-Galmier (Loire).⏎Seven years later, he found a partnership with Bessac and the painter Joséphus Veyrat. Then, he set up his own business in Grigny, around ten kilometres from Lyon. Such
-
home.” –LHBGC, p. 37⏎When Bailey joined the Horticulture faculty at Cornell University in 1888, he immediately set to work expanding that program and also establishing his family in a comfortable home on Cornell
-
to the film. Rather than Australia becoming a frozen tundra, “[we] would be fried,” according to Kusenko.⏎And then, when our heroes set off a stellar bomb successfully in the heart of the sun, it would take much
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 is the print echo print echo on a print echo on a print echo print echo print echo print print print print print print print print print print print print print print print print print print print print print print print print print print print print print
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
head of the State and all executive acts are done in his name. However, the Governor is bound by the aid and advise of the elected State Government under article 163 of the Constitution. It cannot be said that the Governor, who
-
of two molecules with complementary rather than identical structures in juxtaposition. We see here a direct connection to molecular packing in molecular crystals. Aleksandr Isaakovich Kitaigorodsky (left) published a brief paper in 1945
-
do just that. After discussion with another friend, Wes Prichard, Prichard suggested to Wennberg<|endoftext|>Home > Preview⏎The flashcards below were created by user⏎on FreezingBlue Flashcards.⏎What factors could affect feet and
-
the volcano has fumaroles that emit gas in the crater on the main summit of Kibo.⏎Mount Elbrus, Europe’s highest peak stands at 5,642
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
- L20:1511 excites
- L21:2174 weakly suppresses
- L19:2831 weakly excites
- L11:2971 weakly excites
- L21:2022 weakly suppresses
- L8:112 weakly suppresses
- L2:379 weakly suppresses "the" right after "without"
- L5:2642 weakly excites
- L19:2126 weakly suppresses
- L10:706 weakly suppresses
- L8:2100 weakly suppresses
- L1:2440 weakly suppresses
- L20:1793 weakly suppresses
- L22:530 weakly suppresses
- L11:1195 weakly suppresses
- L17:1403 weakly excites
- L2:2183 weakly excites
- L18:1844 weakly suppresses
- L20:579 weakly suppresses
- L15:1850 weakly excites
- L16:1613 weakly suppresses
- L5:1772 weakly suppresses
- L20:2681 weakly suppresses
- L8:2811 weakly excites
Feeds later layers
- L27:1150 suppresses
- L27:1484 excites
- L25:2283 suppresses
- L27:864 weakly suppresses
- L24:2727 weakly excites
- L26:1297 weakly excites
- L24:1173 weakly excites
- L27:1058 weakly excites
- L25:2543 weakly excites
- L25:1726 weakly excites
- L27:850 weakly excites
- L26:822 weakly excites
- L24:708 weakly excites
- L24:1151 weakly excites
- L26:2182 weakly excites
- L26:1369 weakly excites 'Pr' starting capitalised names
- L24:2956 weakly excites
- L25:947 weakly suppresses
- L24:1669 weakly excites
- L24:1009 weakly excites
- L24:1577 weakly excites
- L25:1718 weakly excites
- L26:2005 weakly excites
- L24:2863 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
- off
- Off
- -off
- (off
- off
- about
- Off
- About
- offs
- ers
Suppresses
- console
- println
- println
- Console
- console
- prints