'at' in idioms like 'at work'
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 'at' when it starts a set phrase such as 'at work', 'at least', or 'at the heart of'. Turning it up made the model repeat 'work' and 'least' many times.
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.
-
be and negation. If you ask her, "Give me the ball that is not red," she'll give you the red ball. So negation is at the same level of abstraction, it would seem, as time.⏎How does this
-
of care (signs, interfaces with the electronic medical record, and smart phone apps).⏎Why is this decision support ineffective? Three key facets of human psychology are at least partially responsible: 1. We are bad at knowing when we need help
-
been so pervasive – though certainly degrees have varied in different times and places – that regarding it as coincidental seems farfetched; a structural process seems to be at work.⏎The only force in the known world that provides functional and enduring structure among
-
Mexico and parts of Southern California was barely worth noting compared to the ongoing, major forces that shape planet Earth. In a geologic sense, there are greater forces at work, from the coming and going of ice ages to an ongoing imbalance in the
-
from therapeutic breaches to unthinkable crimes that defy our understanding — can be rehabilitated.⏎That last point – whether empathy can be rehabilitated, or taught, is also at the heart of a fascinating melding of art and science that panelist Alice Fl
-
factor that made her quite atypical was her chosen field of art specialization in sculpture rather than the standard popular domain
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: if I could, with my current to be to be, and with my current to be to be, then... Well, obviously, if I can't, then... Well, obviously, if I can't, then... Well, obviously
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
from conflict to cyclone relief.”⏎Senior emergency response specialist, United Nations Office for the Coordination of Humanitarian Affairs⏎Cyclones Idai and Kenneth that swept
-
Top Program to states that joined Common Core. Federal Race to the Top money is being used to fund the tests for Common Core, which are extensive
-
Nearly 7,000 people have been displaced and more than 5,000 homes destroyed, damaged or flooded,” said the UN Office for the Coordination of Humanitarian Affairs (OCHA), which also warned that the death
-
had been done. It encourages the doer to double up and seek remedy in ways unimaginable. In other words, the feeling of ‘There is light at the end of the tunnel‘ will spur one to do better.⏎3 Knowledgeable —
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
- L19:2538 excites "at" ending a phrase
- L17:1325 excites
- L18:895 suppresses
- L4:430 suppresses
- L8:2038 suppresses
- L21:2846 excites 'at' after a job title, before an institution
- L19:2572 weakly excites
- L13:2374 weakly excites
- L18:2550 weakly suppresses
- L22:3023 weakly suppresses
- L14:707 weakly suppresses
- L0:2837 weakly excites
- L22:1284 weakly excites
- L22:2683 weakly suppresses
- L4:1054 weakly suppresses
- L4:862 weakly suppresses 'to' after 'too big' or 'right angles'
- L6:2979 weakly excites
- L6:2614 weakly suppresses
- L16:1530 weakly suppresses
- L10:2041 weakly excites
- L0:1417 weakly suppresses
- L16:1216 weakly suppresses
- L12:1487 weakly suppresses
- L21:1935 weakly excites
Feeds later layers
- L25:965 excites
- L24:2924 excites
- L27:1598 weakly excites
- L25:770 weakly excites
- L27:1895 weakly excites
- L24:2968 weakly excites
- L24:2870 weakly excites
- L24:2114 weakly excites
- L26:840 weakly excites
- L24:1879 weakly suppresses
- L26:2780 weakly excites
- L24:973 weakly excites
- L24:214 weakly suppresses
- L27:946 weakly excites
- L24:1429 weakly excites
- L24:216 weakly excites
- L25:1212 weakly excites
- L24:428 weakly excites
- L24:1298 weakly suppresses
- L25:1730 weakly excites
- L24:1167 weakly suppresses
- L24:1731 weakly suppresses
- L24:2690 weakly excites
- L26:2705 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
- rest
- least
- Least
- once
- (rest
- 工作
- least
- odds
- work
- night
Suppresses
- ohon
- 决
- 迅速
- boxed
- explode
- activex
- umpy
- jected
- alyze
- umping