unclear: time spans and 'once you've'
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. 4 of Qwen3-0.6B. Its strongest snippets are mixed: they peak on different tokens. 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 a few unrelated words, mostly ones that point to a span of time, like 'period', 'year' and 'time', or the 've in 'once you've'. It switches off hard on 'when' starting a new clause. Turning it up made the model's text drift into random topics, so the steering results don't reveal much.
Written by Claude from this neuron's activations and steering outputs. Confidence: low. 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.
-
"Certain trends are increasing longevity and others are reducing it," he says. "I'm trying to figure out how much each of these trends contributes."⏎Over the period Lichtenberg studied a huge number of
-
of a number, you should first know how many decimal places it falls under. You can do this by multiplying the number by the maximum possible mark. Once you’ve determined the maximum, you can also divide it by the total marks. This will
-
drive to another, we copy from a hard drive to a file. Drive images are more flexible, as you can do what you please with the data once you’ve pulled it off the source drive.⏎Your drive image is going to be a big
-
Let’s look at powder coating vs. galvanizing and when to use one or the other. Or even both!⏎Contact Minnesota Industrial Coatings when you’re ready to apply America’s best industrial coating solutions. We provide our customers with consistently high
-
hardboiled eggs or granola with fruit, nuts and seeds.⏎Eliminate or Min<|endoftext|>Venezuela’s human rights record was assessed under the UN Universal Periodic Review, whose report was adopted in March. Venezuela had accepted several of the
-
0, shows that even before the COVID-19 pandemic, majorities of citizens reported going without needed medical care and clean water at least once during the previous year. Nigerians give the government low marks for its performance in providing water and sanitation
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 am looking for the right to be the time I have to make the decision to choose the right, I am thinking about when the time I have to be the time I have to make the decision to choose the right, I am thinking
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
he was so wise that none could question him concerning anything but that he knew the solution. He went up and down the earth to give instruction to men; and when he came upon invitation to the abode of certain dwarves, Fjalarr
-
to attack the seemingly impregnable fortress of Fort Wagner. The odds were heavily against them; the fort had never been taken. America was watching them – and when the bugle sounded they marched towards the guns, flag held high.<|endoftext|>A
-
care and love for all whom she met—represents the best in Christian witness.⏎The poor white residents of Glynn County also trusted Deaconess Alexander. When the Depression hit the rural poor, she became the agent for government and private aid
-
:83] And when they listen to the revelation received by the Messenger, thou wilt see their eyes<|endoftext|>
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
- L1:2789 suppresses
- L2:2166 suppresses 'in' after high or low
- L2:383 weakly excites
- L2:2451 weakly suppresses
- L0:178 weakly suppresses
- L1:1681 weakly excites
- L0:1301 weakly suppresses
- L0:1224 weakly suppresses
- L2:49 weakly suppresses
- L2:111 weakly suppresses
- L0:1405 weakly suppresses
- L1:1464 weakly excites
- L0:2920 weakly suppresses
- L1:886 weakly excites
- L2:537 weakly excites
- L0:588 weakly suppresses
- L0:2018 weakly suppresses
- L2:75 weakly suppresses
- L2:1710 weakly suppresses
- L3:691 weakly excites
- L1:2851 weakly excites
- L2:766 weakly excites
- L0:3009 weakly excites
- L1:461 weakly excites
Feeds later layers
- L10:833 suppresses
- L5:55 suppresses
- L8:1749 suppresses
- L13:2439 suppresses
- L17:1662 suppresses
- L10:456 suppresses
- L18:3002 suppresses
- L16:583 suppresses
- L15:2342 suppresses
- L17:619 suppresses
- L16:2083 suppresses comma after 'in this article'
- L7:451 suppresses
- L9:306 suppresses
- L7:1484 weakly suppresses
- L5:892 weakly suppresses
- L19:2872 weakly suppresses
- L8:691 weakly suppresses
- L16:2726 weakly suppresses
- L7:1182 weakly suppresses
- L17:2741 weakly excites
- L6:2369 weakly suppresses
- L15:513 weakly suppresses
- L8:1916 weakly suppresses
- L9:811 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
- iveness
- 渚
- скую
- REW
- insula
- ราม
- acity
- ifferential
- ucer
- opping
Suppresses
- 谈到
- 的时候
- entimes
- .ComponentModel
- indebted
- 大宗
- 谈起
- ึก
- meld
- (Token