the word 'figure' in 'figure out'
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 lights up on the word 'figure' (or 'figured' and 'figuring') when it is followed by 'out', as in 'figure out what's causing it'. When it was turned up, the model got stuck repeating the word 'out' over and over.
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.
-
take steps to clear up the haze after you figure out what’s causing it.⏎What Is Brain Fog?⏎Brain fog is not a
-
correct answers. Victor doesn’t know any answer, but is allowed to take the same test several times. Can Victor work out a strategy that guarantees that he can figure out all the answers after the 29th attempt? after the 2
-
it and how it threatens me? "- These questions sound in the office of a therapist and are driven into the search box of the browser.Let's try to figure it out.⏎What kind of substance: urea⏎Chemically it is a
-
person will make different motions, and whenever s/he changes the motion, everyone will follow. When the person who left the room returns, s/he will try to figure out who the leader is.⏎A
-
How to grow a reader⏎Tips for turning a fledgling reader into a voracious one.⏎By Peg Tyre⏎Once your child has figured out how to decode words and can actually read in a sustained way, then a
-
our genes surprised us,’ says study coordinator Tuuli Lappalainen, previously at UNIGE and now at Stanford University. ‘It is important that we figure out the general laws of how the human genome works, rather than just delving
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: I can be a great hero to be a good hero, and be a good hero to be a good hero. Each time I can be a good hero, I can be a good hero to be a good hero to be a good hero
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
would come to an end."⏎"The French empire was one of the custodians of the⏎Roman Catholic faith, and, like the Italian empire, supported
-
physics has been made at University of Konstanz. The team of scientists headed by...⏎Yersiniae cause severe intestinal infections. Studies using Yersinia pseudotuberculosis as a model organism aim to elucidate the infection mechanisms of these
-
28)⏎Design & Thinking is a documentary released this year about design thinking. I haven't seen it yet, but I'm loving the approach -- the cacophony of voices
-
isation.⏎The fundamental challenge for a young person now is how does one live with oneself, and how does one live with others in a world filled with a cacophony of voices, with
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
- L20:533 excites
- L20:1550 excites 'find' in 'to find out'
- L17:241 excites
- L4:1366 weakly excites
- L16:1672 weakly excites
- L18:2542 weakly suppresses
- L22:288 weakly suppresses
- L8:262 weakly suppresses
- L20:78 weakly suppresses
- L19:107 weakly excites
- L20:256 weakly suppresses
- L0:785 weakly excites
- L22:2955 weakly excites
- L19:1287 weakly suppresses
- L20:2709 weakly excites
- L7:1918 weakly suppresses
- L1:1152 weakly excites
- L3:2434 weakly suppresses
- L17:577 weakly suppresses
- L14:311 weakly excites
- L4:2692 weakly excites
- L2:2821 weakly suppresses
- L21:443 weakly suppresses
- L21:305 weakly suppresses
Feeds later layers
- L27:850 excites
- L26:1474 excites
- L27:1965 weakly excites
- L27:1301 weakly excites
- L24:2967 weakly excites
- L25:1827 weakly suppresses
- L24:2884 weakly suppresses
- L25:2124 weakly excites
- L24:2899 weakly suppresses
- L26:1577 weakly suppresses
- L25:1453 weakly suppresses
- L24:872 weakly excites
- L25:2021 weakly excites
- L25:2030 weakly suppresses
- L27:2486 weakly excites
- L24:2567 weakly suppresses
- L24:2822 weakly suppresses
- L24:2236 weakly excites
- L25:333 weakly suppresses
- L27:2705 weakly suppresses
- L25:1541 weakly suppresses
- L24:2495 weakly excites
- L24:1513 weakly suppresses
- L25:2650 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
- out
- -out
- out
- 出
- (out
- outs
- →out
- Out
- _out
- Out
Suppresses
- nist
- ills
- olio
- su
- ioni
- creds
- cem
- tał
- inated
- ione