the word 'still' mid-sentence
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 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
Fires on the word "still" when it sits between a subject and its verb, as in "has still been unable" or "is still evolving." Turning it up did not produce any obvious shared theme in the output.
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.
-
seizures and that one of the students had previously suffered seizures.⏎But Susan Rafla, the mother of the other child who suffered a seizure, said she has still been unable to "rule in or rule out" the possibility that the portable building
-
d'Anticosti and Îles de la Madeleine), Nova Scotia (including present-day N.B. and P.E.I.). Hudson's Bay Co. still administers Rupert's Land. Louisiana is ceded to Spain by France.⏎Following
-
manual tools—so that if one kind fails, another can intercede.⏎If everything goes wrong, and your reactor melts down, don't give up. You still have evacuation<|endoftext|>Mohammad H
-
the infection, suggests John Jernigan, MD, director of the CDC's office of health care infection prevention, research, and evaluation.⏎"The attack rate is still to be determined, but
-
De Stael, a French author, at one time invested in the town lands, on the advice of her friend Gouverneur Morris, while it was still part of Pierrepont.⏎A history in 1894 indicated that
-
being gay in some parts of the country. As a result, LGBTQ persons in the United States may still face some challenges not experienced by non-LGBTQ residents, particularly in the Deep
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 a robot that is not for the public." and then is the example sentence. The first thing is the thing that is not the first thing, and the last thing is the last thing. The first paragraph is about the first
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
proportional to the change in concentration (C, higher concentration less lower concentration) divided by the thickness of the membrane. At steady state, Fickian diffusion can be approximated by:⏎J = D (C
-
-sustaining arc across electrodes placed 5 mm apart. A little math told him that he
-
match experimental results found in the literature. These results are for steady-state solutions; a dynamic implementation requires transient modeling as well. This can be achieved by
-
to staying up-to-date, boosting your immune system is wellness 101. Better yet
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
- L3:2720 weakly suppresses
- L0:1751 weakly suppresses
- L3:115 weakly suppresses
- L0:1885 weakly excites
- L2:41 weakly suppresses
- L3:2169 weakly excites
- L3:2071 weakly excites
- L0:2953 weakly excites
- L2:2810 weakly excites
- L1:2252 weakly suppresses
- L3:550 weakly suppresses
- L1:1674 weakly excites
- L2:803 weakly suppresses
- L1:1239 weakly excites
- L3:1934 weakly excites
- L2:2574 weakly suppresses
- L3:1709 weakly suppresses
- L1:2408 weakly excites
- L2:1888 weakly excites
- L2:111 weakly suppresses
- L2:1760 weakly suppresses
- L1:2917 weakly excites
- L2:2624 weakly excites
- L1:1440 weakly suppresses
Feeds later layers
- L7:1079 excites
- L17:2158 excites
- L16:2123 excites
- L7:1884 excites
- L10:696 excites
- L13:2890 excites "still" or "as long as" meaning continuing
- L11:1853 excites
- L9:1699 excites
- L18:590 excites
- L16:937 suppresses
- L13:1147 suppresses
- L15:1345 excites
- L20:2841 weakly excites
- L8:2072 weakly excites
- L13:2523 weakly suppresses
- L5:358 weakly excites
- L15:2779 weakly excites
- L9:1892 weakly excites
- L26:540 weakly excites
- L14:2743 weakly excites
- L14:1874 weakly excites
- L5:2146 weakly excites
- L8:2486 weakly excites
- L19:1820 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
- LAG
- `"
- `'
- pys
- AINS
- OG
- casts
- esy
- blo
- ไ
Suppresses
- soever
- nex
- uç
- uthor
- SPECIAL
- radi
- ALLE
- rush
- 答卷
- entral