'on' in 'went on to'
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. 5 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
The neuron fires on the word 'on' in phrases like 'went on to' or 'go on to', where someone later does something. Turning it up did not clearly change the writing style.
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.
-
and small, in the sciences as well as in the humanities.⏎Attending a lecture and taking notes works very well for some students, particularly the students who go on to careers in academia. They’re able to assimilate and process the information shared
-
to Alzheimer’s in roughly 10% to 15% of cases each year, according to the study.⏎Forty-five percent of the participants went on to develop serious dementia consistent with Alzheimer’s disease. Those who had high cholesterol,
-
and Technology (NIST) and replaced by the Advanced Encryption Standard (AES), Diebold went on to hardcode it into the source code—a willful circumvention of best practices
-
This mansion in the Vine neighborhood, built in 1910, was the home of Alonzo Herndon, who was born into slavery but went on to become Atlanta's first black millionaire. (Photo Courtesy by Philip Graitcer
-
Schultz, PhD, director of the Center for Autism Research at CHOP and a co-author on the study.⏎These findings suggest that 7-month-olds who go on to develop autism show subtle, yet overt, behavioral differences prior to the emergence of
-
action for equality. The first IWD was acknowledged in the U.S. in 1909 when 15,000 female workers went on strike in New York City to demand improved working conditions, shorter working hours and voting
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:** ##** A** All ` ##** A** A** A** A** A** A** A** A** A** A** A** A** A** A** A** A** A** A** A**
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
times the mass of the sun was converted to energy in a fraction of a second.⏎The discovery has been accepted for publication in Physical Review Letters.⏎Gravitational waves are ripples in the universe caused by some of the most energetic cosmic events,
-
ipples started to get serious.⏎But the first generation LIGO experiment, which ran for eight years, wasn’t sensitive enough. Which is understandable. Gravitational waves are minuscule — the atomic jitters that pass through our world when two
-
modo have been flying for a month, the LIGO team wanted to be absolutely certain before making an official announcement.⏎That announcement has just come. Gravitational waves were observed on 14 September 2015, at 5
-
a century ago, physicists have been on the hunt for gravitational waves, ripples in the fabric of spacetime. That hunt is now over. Gravitational waves exist, and we’ve found them.⏎That’s according to researchers at the Laser
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
- L3:2935 weakly suppresses
- L4:1724 weakly suppresses
- L0:1085 weakly excites
- L1:2313 weakly suppresses
- L4:2082 weakly suppresses
- L0:1572 weakly excites
- L4:1684 weakly suppresses
- L4:1566 weakly excites
- L4:2477 weakly suppresses sorting and ordering words
- L3:3039 weakly suppresses
- L4:1602 weakly excites
- L4:2567 weakly suppresses the word 'still' mid-sentence
- L4:2794 weakly excites
- L3:1472 weakly suppresses
- L4:1846 weakly excites
- L3:781 weakly excites
- L4:2367 weakly suppresses
- L4:738 weakly suppresses
- L1:2540 weakly excites
- L2:2396 weakly suppresses
- L4:2860 weakly suppresses
- L4:2652 weakly suppresses
- L3:810 weakly excites
- L2:2570 weakly suppresses
Feeds later layers
- L24:2031 weakly suppresses
- L7:1748 weakly excites
- L19:1233 weakly excites
- L9:1243 weakly suppresses
- L8:1798 weakly suppresses
- L8:795 weakly excites
- L11:1228 weakly excites
- L11:271 weakly excites
- L11:1103 weakly excites
- L12:2170 weakly suppresses
- L8:913 weakly suppresses
- L16:2872 weakly excites
- L11:604 weakly suppresses
- L16:1154 weakly excites
- L25:663 weakly excites
- L11:3024 weakly suppresses
- L10:2533 weakly excites
- L10:1539 weakly excites
- L8:1779 weakly excites
- L23:550 weakly suppresses
- L10:1447 weakly suppresses
- L21:99 weakly suppresses
- L8:3065 weakly excites
- L22:2200 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
- -sc
- agination
- setDescription
- forward
- entieth
- olio
- survived
- -setup
- -over
- 毕竟是
Suppresses
- 使人
- .poi
- selon
- 怠
- derail
- stead
- reet
- nameof
- mü
- 使我