'from' after 'Retrieved' and source phrases
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. 5 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 'from' in citation phrases like 'Retrieved from', and in phrases like 'can be purchased from' or 'expected too much from'. It goes negative on 'by' after passive verbs such as 'acquired by' or 'bought by'. Turning it up or down did not produce a clear change in the writing.
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.
-
of Publication⏎Blanks, C. M. (1948). San Antonio the Inland Gate. Retrieved from https://digitalcommons.pvamu.edu/pvamu-theses/39
-
eating disorders. The trait of perfectionism is common among people with obsessive-compulsive disorders. Or it could be a response to having parents who expected too much from you. Maybe they never let you off the hook, even if you got
-
6C>.⏎- APA: Learning Math; Measurement: Indirect Measurement And Trigonometry. Boston, MA: WGBH Media Library & Archives. Retrieved from http://openvault.wgbh.org/catalog/V_4E4EE3
-
although some states also award goods or services. In many cases, a winner must pay taxes on the prize amount before receiving it.⏎Lottery tickets can be purchased from retail stores, newsstands, gas stations, and convenience stores. Some lotter
-
1853-1863). International Journal on Stereo & Immersive Media, 2(2), 54-69. Retrieved from https://revistas.ulusofona.pt/index.php/stereo/article/view/
-
]. (2018, July 6). 1929 Stock Market Crash and the Great Depression –⏎Documentary [Video File]. Retrieved from https://youtu.be/qlSxPouPCIM⏎The post
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: there is a 200-gram mass of 10 grams. The mass of the object is directly proportional to the product of the kinetic energy, which is 200 grams. The product of the kinetic energy is
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
is closed to the public. It is undergoing a state-ordered cleanup for soil and groundwater contamination linked to Peerless Photo Products, a previous owner that was acquired by Agfa.⏎A security guard who never heard of Nikola Tesla turned away visitors on
-
sold to a private owner.⏎Over the years, the hotel changed hands several times and underwent several renovations. In 2007, the hotel was purchased by the French Lick Resort and underwent a $500 million restoration, restoring
-
is not a teacher. Coleman worked for McGraw-Hill (textbook publishing company) before leaving to create GROW a curriculum testing company which was later bought by McGraw-Hill. Bill Gates (who funded the creation of Common Core)
-
, a tag library defines a set of Java classes for the custom JSP actions that the tag library supports. Invoking the tag library imports custom processing and actions into the Java environment so a Java runtime environment like JRun 3.0 will
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
- L3:822 suppresses
- L4:1038 weakly excites
- L0:2358 weakly suppresses
- L0:1923 weakly suppresses
- L4:97 weakly suppresses
- L3:3055 weakly excites
- L2:2661 weakly suppresses
- L2:2895 weakly suppresses
- L4:2050 weakly excites
- L2:1352 weakly suppresses
- L3:1568 weakly suppresses
- L4:557 weakly suppresses
- L3:2005 weakly excites
- L1:2846 weakly excites
- L3:1694 weakly suppresses
- L4:1662 weakly suppresses
- L4:1528 weakly suppresses
- L4:2325 weakly suppresses
- L0:1797 weakly excites
- L3:1342 weakly excites
- L0:1839 weakly excites
- L4:12 weakly suppresses
- L1:1720 weakly suppresses
- L3:2375 weakly suppresses
Feeds later layers
- L9:2267 excites 'with' after exchange nouns
- L21:2262 weakly excites
- L9:654 weakly excites
- L17:2984 weakly excites blank before a number in prevalence claims
- L22:2344 weakly excites number after 'account for' or 'over'
- L15:2416 weakly suppresses
- L7:1393 weakly excites
- L21:1203 weakly excites
- L6:2695 weakly suppresses
- L19:2244 weakly excites
- L17:978 weakly excites
- L7:1982 weakly suppresses
- L25:1488 weakly excites
- L14:725 weakly excites
- L8:1034 weakly excites
- L27:2847 weakly excites
- L9:2461 weakly excites
- L6:1834 weakly excites
- L17:2711 weakly excites
- L18:3001 weakly excites
- L9:186 weakly suppresses
- L7:1028 weakly excites
- L15:2463 weakly suppresses
- L7:1326 weakly excites
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
- mitted
- 用户提供
- .Generation
- 屏幕上
- 更多信息
- 媒体报道
- 官方微信
- 层出
- nią
- 覆盖率
Suppresses
- uld
- 降落
- oard
- dress
- 竿
- adh
- dsl
- scoped
- ë
- rious