'without' introducing a contrast
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. 17 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 'without' when it links two ideas, like 'reduce the temperature without generating steam'. When it was turned up, the model kept repeating words like 'being', so the effect is loose.
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.
-
to balance this equaliser? You can put more⏎than one weight on a hook.⏎What is the greatest number of counters you can place on the grid below without four of them lying at the corners of a square?⏎How many different triangles can
-
marked transformation with the discovery of an aspect of enzyme evolution called "moonlighting." That is, a protein can acquire a second (moonlighting) function without concomitantly losing all or part of its original function
-
–though not excessive–application of inorganic fertilizer boosted rice yields and reduced land use and water use, all without substantially increasing greenhouse gas emissions. The findings also showed that high-yield farming often
-
, a quick two-to-three-second dash with a straight stream will reduce the temperature without generating a massive amount of steam.3⏎The direct attack method is designed to
-
values in the IO or (ST s) monads, if you need them.⏎> That is, I can't change a⏎> component of somethingNew without creating y. The point is there's⏎> nothing about x that changed,⏎
-
in this woman or her story. But then she gets to Africa and everything shifts. Caitlin must find a way to keep elephants out of the farmers’ crops without
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 number of, as a whole, of, and, and, and, and, and, and, and, and, and, and, and, and, and, and, and, and, and, and, and,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
, historical analysis, and any understanding of broader systemic relations. In this regard, it does the opposite of critical memory work by eliminating those public spheres where people learn to translate private troubles into public issues. That
-
knows more about his or her learning strategy than anyone else. So if you'd like to learn something about the best strategy to use when teaching your students to fly, why not just ask them to tell you?⏎The next time you're
-
that changes how the⏎cd command works. Normally when you try and cd⏎into a directory, it will only try and change to the path specified
-
the free Adobe Acrobat Reader.)⏎Click for the answers to this week's puzzle.⏎Click for Sudoku Puzzles from previous weeks.⏎Click for tips for teaching students to do Sudoku puzzles.⏎Copyright © 2006 Education World<|endoftext|>1.Model
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
- L16:797 suppresses
- L6:2882 suppresses
- L13:2681 excites
- L16:1153 suppresses
- L16:513 suppresses
- L6:2064 suppresses
- L6:2747 suppresses
- L14:526 excites
- L6:1133 excites "to" before an infinitive verb
- L1:943 weakly suppresses
- L9:226 weakly suppresses 'to' starting an infinitive
- L15:136 weakly suppresses
- L10:707 weakly excites
- L15:210 weakly excites
- L14:1091 weakly excites
- L15:1186 weakly excites
- L15:27 weakly excites
- L11:223 weakly suppresses
- L16:314 weakly excites
- L13:44 weakly excites inside web addresses and citations
- L11:1300 weakly excites
- L5:763 weakly suppresses
- L6:2407 weakly suppresses
- L8:1703 weakly suppresses
Feeds later layers
- L24:140 strongly excites
- L22:1425 strongly excites
- L22:609 strongly excites
- L26:686 strongly excites
- L21:1441 strongly excites
- L23:424 excites
- L25:620 excites
- L27:410 excites
- L25:376 excites
- L27:1237 excites
- L26:252 excites
- L27:1890 excites
- L25:218 excites
- L27:677 excites
- L26:708 excites
- L27:1148 excites
- L21:1521 suppresses
- L26:808 excites
- L27:1545 excites
- L27:303 excites
- L25:1596 excites
- L26:1304 excites
- L24:1811 excites
- L26:733 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
- being
- needing
- having
- being
- requiring
- providing
- having
- étant
- allowing
- preferring
Suppresses
- ichten
- iless
- 我会
- oster
- arose
- "]];⏎
- ']];⏎
- by
- saja
- veillance