'turn' in 'which in turn'
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. 2 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 'turn' in the phrase 'which in turn', which links a cause to its knock-on effect. Turning it up did not make the model's text visibly use this phrase.
Written by Claude from this neuron's activations and steering outputs. Confidence: medium. 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.
-
. The trees that die as a result of rising water levels attract insects, which in turn feed woodpeckers, whose holes later provide homes for other wildlife.⏎Figure
-
production will be the most affected by climate change by 2030. This would lead to food insecurity, which in turn could lead to social instability, as was already the case in Mozambique in
-
film, including one in which the dwarfs build and carve<|endoftext|>Gothic architecture is a good architectural design which in turn was especially popular in Europe throughout the Middle Age range, lasting into the middle and
-
onions, have been shown to reduce inflammation in the body, which in turn lowers the risk of artery hardening. Fresh, peeled beets can be thinly
-
secure seal that prevents particle, vapor, and fluid leakages, which in turn protects and maintains pressure whilst keeping debris out, ensuring proper function and safety.<|endoftext|>
-
initial state of the atmosphere is critical. Even if Doppler radar information may seem highly localized, we find that it enhances the regional atmospheric conditions, which, in turn, can significantly improve the dynamic prediction of how the monsoon depression will move as
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: the time that, with a few years' passage, to the equality of two variables... I think the argument of reverse the inequality, I am not so much the time as... and so. that the previous and subsequent, and equal,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
in one of its caves during the Arabic period. This tradition is reflected in the writings of Fray Antonio Agustín de Milla and Suazo, dated back to the 17th
-
, America embraces people of all faiths and of no faith. We are Christians and Jews, Muslims and Hindus, Buddhists and Sikhs, atheists and agnostics. Our religious diversity enriches our cultural fabric and reminds us that what binds
-
values - before there developed⏎the actual beginnings of the Philippine revolution that fell under the class leadership of the ilustrados or the liberal bourgeoisie⏎that surrounded Aguinaldo.⏎In this Propaganda Movement, Dr. Jose Rizal made
-
third of U.S. adults under age 30 are unaffiliated.⏎By the year 2050, the study projects, atheists, agnostics and religiously<|endoftext|>Sea Monster Eggs and Leadership⏎In the early 9
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
- L0:2204 weakly excites
- L1:2968 weakly excites
- L1:14 weakly suppresses
- L0:2460 weakly suppresses
- L0:2712 weakly excites
- L0:49 weakly suppresses
- L1:6 weakly excites
- L1:164 weakly suppresses
- L0:1171 weakly excites
- L1:2253 weakly suppresses
- L0:2942 weakly suppresses
- L1:1699 weakly suppresses
- L1:2118 weakly excites
- L0:397 weakly suppresses
- L1:1866 weakly suppresses
- L1:2651 weakly excites
- L1:2357 weakly excites
- L0:1289 weakly suppresses
- L1:486 weakly excites
- L1:1583 weakly excites
- L1:1719 weakly excites
- L0:2841 weakly suppresses
- L1:2307 weakly suppresses
- L0:2099 weakly suppresses
Feeds later layers
- L10:2439 weakly excites
- L5:1789 weakly suppresses
- L4:1154 weakly excites
- L14:2050 weakly suppresses
- L16:2011 weakly excites
- L25:1922 weakly suppresses
- L18:899 weakly suppresses
- L4:646 weakly suppresses
- L27:2023 weakly suppresses
- L5:2330 weakly excites
- L21:369 weakly excites
- L16:606 weakly suppresses
- L16:2102 weakly suppresses
- L16:1644 weakly suppresses
- L17:1165 weakly excites
- L9:1237 weakly excites
- L22:36 weakly excites
- L7:2221 weakly suppresses
- L13:532 weakly suppresses
- L11:1910 weakly excites
- L11:1203 weakly excites
- L19:2410 weakly suppresses
- L13:2579 weakly suppresses
- L6:2423 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
- кат
- 覚
- nable
- bolt
- worthy
- .getHeader
- 檎
- ñe
- =((
- ient
Suppresses
- ursively
- DATES
- udios
- lexible
- 相见
- 羽毛
- psy
- етод
- inality
- arma