the word 'long' in phrases
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. 18 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 'long' in set phrases like 'a long way', 'long ago', 'long before' and 'long-haired'. Turning it up made the model write more about long periods of time, so it seems tied to the idea of duration.
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.
-
of the pelvic floor and the muscles that support the bladder neck and vagina and give these organs tone and strength. A strong healthy set of pelvic floor muscles goes a long way to preventing urinary stress incontinence and pelvic floor prolapse.⏎There is
-
still farmland back then, a good six miles north of teeming downtown and this was long before public transit. Maps of New York City as late as 184
-
this important message!<|endoftext|>Forty-eight years ago, an event occurred that shattered the American psyche and permanently altered society as we know it.⏎The Beatles, a long-haired quartet from Liverpool, England, appeared on the Ed Sullivan show in New
-
dwelling houses and field systems. This was in days when there were still wolves in the British Isles, long before the Romans introduced rabbits and horse chestnut trees and in a time when there
-
more premature deaths each year than automobile accidents, which was long considered the leading cause. According to the study, particulate air pollution causes at
-
necessary context for understanding. But it’s the strange bits of information that can really capture a person’s curiosity and help us relate to the people that lived long ago. They’re also the most fun things to share with our guests and get
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 want to go to a “first step” level, which is the minimum level of “full” “swaption” offering, which is the first “step” of “P0” or “P0” level. But I
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
order to resolve the dispute without the use of force. The rebels, who hail from the autonomous island province of Sulu in the southwestern Philippines, had been given until midnight on Tuesday
-
), Part 4, Chapter 7, eBook.<|endoftext|>Transcript for A New Horizon, segment 05 of 13⏎The Eastern
-
an infected tick. Discover the latest research on babesiosis here.⏎Transcriptional analysis of p30 major outer membrane protein genes of Ehrlichia
-
1–2. Sign up to view the full content.⏎This preview has intentionally blurred sections. Sign up to view the full version.View Full Document⏎Unformatted text preview: 60 | | | | CHAPTER 1 FUNCTIONS AND MODELS
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
- L10:852 strongly excites
- L4:1978 excites
- L5:2290 excites
- L17:1608 excites
- L16:155 excites
- L8:918 excites
- L17:509 suppresses
- L10:273 suppresses
- L12:1703 excites
- L5:2731 weakly suppresses
- L17:378 weakly suppresses
- L10:937 weakly suppresses
- L1:2130 weakly suppresses
- L7:1101 weakly suppresses
- L1:27 weakly excites
- L2:2166 weakly suppresses 'in' after high or low
- L9:515 weakly suppresses
- L6:3028 weakly suppresses
- L10:1502 weakly excites
- L17:1445 weakly excites
- L17:725 weakly excites
- L4:2251 weakly excites
- L9:2572 weakly suppresses
- L5:1093 weakly excites
Feeds later layers
- L19:1712 strongly excites
- L22:2698 excites 'long' in 'in the long run'
- L24:712 excites 'long' in 'long-term' and 'long-lasting'
- L20:2982 excites
- L27:1469 excites
- L22:1699 excites
- L22:1465 excites
- L19:1768 suppresses
- L19:2901 excites
- L20:2368 weakly suppresses
- L20:1736 weakly excites
- L25:2988 weakly excites
- L19:2850 weakly excites
- L23:1490 weakly excites
- L22:1915 weakly suppresses
- L21:2289 weakly excites
- L21:2818 weakly excites
- L22:2755 weakly suppresses
- L20:1163 weakly suppresses
- L22:622 weakly excites
- L27:2815 weakly suppresses
- L24:1809 weakly excites
- L26:2136 weakly excites
- L22:1893 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
- -lasting
- dài
- 很长
- lasting
- lasts
- lengths
- lasted
- 最长
- length
- طول
Suppresses
- %\
- #{
- irates
- matter
- -degree
- ula
- rade
- gression
- itals
- >}</