unclear: link fragments and document ends
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. 13 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 is hard to pin down. Its strongest moments come inside web links, such as the middle of a YouTube or Twitter address, and right where one web page ends and another begins. The steering outputs do not show a clear link to these spots.
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.
-
et paper. At a much less sophisticated level, PutnamPig’s “Swine Flu Song” on Youtube http://www.youtube.com/watch?v=tbt_PuVAVTU defends pigs against unfair stigmatisation over H1N1
-
8, July 6). 1929 Stock Market Crash and the Great Depression –⏎Documentary [Video File]. Retrieved from https://youtu.be/qlSxPouPCIM⏎The post 1929 Stock
-
To give a sense of the true dimensions of the Larsen C iceberg, here's an animation based on real data, with something familiar for scale pic.twitter.com/OdQA07qhBb⏎— Adrian Luckman (@adrian_l
-
I highly recommend reading Bruce Rogers' post:⏎Good study for the Moxy people. I did a post and review on that article. https://t.co/D9NIorgXuc— Bruce Rogers (@bjrmd) August 30
-
geographic regions.(opens in new tab)⏎Journal of Clinical Microbiology⏎Suleyman FelekY Rikihisa⏎George RidingPeter Willadsen<|endoftext|>Today, various types of compressed gases such as acetylene, propane, oxygen
-
014⏎Originally published: August 18, 2014. Last Updated: August 18, 2014.<|endoftext|>The more I read about this, the more confused I become. Illuminance,
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 have a 3D space, and the next step is to go to the next step, and the next step is to go to 4th step, and the next step is to go to 5th step, and the
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
) hypotonic D) none of these I think it's B. I'm not too sure though.⏎November 27, 2007<|endoftext|>Hopefully your child will not experience at first hand the impact of war. But sadly
-
oviridae family. …⏎- Noroviruses. …⏎- Pasteurella. …⏎- Yersinia enterocolitica.<|endoftext|>LACMA May Fund The Fading Watts Towers⏎Due to a severe budget
-
ycetica<|endoftext|>Adding Fractions Level 3⏎The 5-step planIn Adding Fractions
-
99) 23: 25. doi:10.1023/A:1005462530658⏎- 896 Downloads⏎Using recently available data from China
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
- L10:2913 suppresses
- L5:3040 weakly suppresses
- L9:1028 weakly excites
- L10:57 weakly suppresses
- L6:2839 weakly suppresses
- L12:1721 weakly excites
- L6:2556 weakly suppresses
- L12:1877 weakly suppresses
- L10:2001 weakly excites
- L10:2621 weakly excites
- L2:2740 weakly suppresses 'on' before time or track
- L8:10 weakly excites
- L8:1837 weakly suppresses
- L1:51 weakly excites
- L4:2760 weakly excites
- L12:1664 weakly excites end of document token
- L12:59 weakly suppresses
- L3:3011 weakly suppresses
- L5:198 weakly excites unclear: weak "and" in lists
- L12:2472 weakly excites
- L11:1539 weakly excites
- L6:214 weakly suppresses
- L6:1710 weakly excites
- L8:2395 weakly suppresses
Feeds later layers
- L19:738 excites
- L18:921 weakly excites
- L25:1816 weakly excites
- L17:1571 weakly suppresses
- L14:2391 weakly suppresses
- L16:1451 weakly suppresses
- L16:1161 weakly suppresses
- L27:877 weakly excites
- L17:1175 weakly suppresses
- L14:471 weakly suppresses
- L18:1723 weakly excites
- L20:2754 weakly suppresses
- L26:135 weakly excites
- L24:782 weakly excites
- L21:2125 weakly suppresses
- L24:483 weakly suppresses
- L17:1252 weakly suppresses
- L14:118 weakly suppresses
- L16:2904 weakly excites
- L25:2846 weakly excites
- L27:1023 weakly excites
- L14:2291 weakly suppresses
- L18:2002 weakly suppresses
- L18:672 weakly excites zero right after a decimal point
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
- 钋
- aye
- illas
- weg
- 体制
- ás
- encoding
- 世家
- dq
- qw
Suppresses
- �
- iterate
- leting
- $is
- rement
- оч
- ﷺ
- ometric
- -second
- ходит