word before an evaluative adjective in essays
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. 14 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
It fires on the word just before a judgement word in generic, polished essay writing, such as 'crucial', 'significant', 'pressing' or 'extended'. Turning it up made the model produce loops of 'from' and 'and', and turning it down produced nonsense words.
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.
-
In the next few years we will see -mandated by legislation in 200
-
In the South, the Rocky Mountains are characterised by thick wood and used to be inhabited
-
can breathe both underwater and on land. While they require air to survive, they have unique physiological adaptations that enable them to hold their breath for extended periods. However, turtles cannot survive indefinitely underwater and must resurface to breathe.⏎
-
between 2027 and 2028.⏎In 2020, the next-generation spacecraft underwent its initial test flight, marking a significant milestone in China's pursuit of space exploration. Following the successful operation of China's
-
and Pigment Granules⏎Have you ever wondered what makes up the structure of your hair? The intricate anatomy of our hair strands comprises several distinct components, each playing a unique role in the appearance and health of our hair. In this article,
-
challenge of optimizing app performance, ensuring compatibility across multiple devices, and addressing varying screen sizes and resolutions. Additionally, staying up-to-date with emerging technologies and frameworks is crucial to deliver cutting-edge apps that cater to users’ evolving expectations. Moreover, debugging
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: Fororetical forthal forthal forthal for forthal forstraints forthal fortheses forchemes for Loretical fororetical forцип forresults forthal forftware forftware forцип forftware for伦限限限伦限限chemes for
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
François Mitterand who, in 1954, defiantly pronounced that ‘L’Alg
-
From the Popul Vuh (Popol Wuj in modern spelling), the Mayan equivalent of the Bible, which states that humans were
-
due to the Aizoaceae (Ice Plant Family), Asteraceae (Daisy family) and Campanulaceae (
-
and of the northern Yangtze River delta - including the Jiangsu province - , as well as in the Fenwei plains area of Shanxi (Shaanxi and Henan provinces).⏎Smog caused by the pollution can result in poor visibility
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
- L10:80 strongly excites
- L12:1091 strongly suppresses
- L9:23 strongly suppresses
- L13:1255 strongly excites
- L8:642 strongly suppresses
- L10:58 excites
- L13:40 suppresses
- L8:609 suppresses
- L10:222 suppresses
- L8:1104 suppresses
- L3:17 suppresses
- L3:157 excites
- L13:2018 excites
- L11:533 suppresses
- L9:302 excites
- L7:276 suppresses
- L8:2902 excites
- L13:3041 excites
- L4:121 excites
- L8:2083 suppresses
- L12:1696 suppresses
- L8:1415 suppresses
- L8:663 suppresses
- L11:2500 suppresses
Feeds later layers
- L16:926 strongly excites
- L16:798 strongly suppresses
- L18:232 excites
- L15:2979 suppresses
- L15:2706 suppresses
- L15:1272 suppresses
- L18:364 suppresses
- L16:704 suppresses
- L20:326 suppresses
- L18:1633 suppresses
- L19:303 suppresses
- L15:1178 excites
- L17:283 excites
- L15:1317 excites
- L16:2610 suppresses
- L17:551 suppresses
- L16:1543 excites
- L15:562 suppresses
- L15:23 suppresses
- L23:99 excites
- L15:910 suppresses
- L17:61 excites
- L19:1466 suppresses
- L17:349 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
- 做
- 做了
- 刚
- Lik
- 3
- [*
- Yours
- Bew
- your
- Bec
Suppresses
- ossal
- baugh
- */)⏎
- \Catalog
- \")
- iameter
- @update
- @⏎⏎
- _circle
- ?")⏎