'relevant' before a noun
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. 8 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 "relevant" (and similar words like "matter" or "topic") when it comes right before the thing being talked about, like "the relevant qualifications". When it was turned up, the writing drifted into unrelated new topics, and turning it down made the text start rambling about a single idea.
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.
-
towards promoting the sport.⏎Swimming education is available in the USA. The department of Education is the largest employer of swimming instructors. Every instructor has to have the relevant qualifications to get employment in the program. The instructors also must have a valid National
-
Yet, Meier says, “I wasn’t too sure about the psychological effects. I hadn’t seen any research on the matter."⏎At the time, Meier was a graduate student, searching for a subject
-
transport make their fitness for transport questionable from a scientific point of view. In a review, the scientific knowledge about fitness for transport is reviewed against the background of the relevant European legislation.⏎Injury and unnecessary suffering⏎The European regulation (Council Regulation
-
activities at home and abroad, the Center for Combating Desertification in Arid and Semi-arid Areas (CCDASA) was launched recently to concentrate on relevant scientific researches on combating desertification supported by
-
developments in machine learning. Probability theory was employed to offer theoretical explanations as to why many machine learning models work (read Murphy’s book for an excellent book on the topic). Support vector machines offered an alternative to neural networks, based on learning
-
, and as a result, sends information directly to the brain via the production of neurotransmitters. The Shanghai University team therefore conducted a systematic review of 21 relevant studies to try and find out more about the potential effects of regulating the microbiota
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 more you think about it, the less you can see, the something in you. It's a process of thinking, which is the first, and the second, and third, and so on. But the something in you is a
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
. Newer standards take a performance or program approach, identifying a safety or health objective and providing a framework for employers to use in achieving that goal. OSHA standards cover the gamut of workplace safety and health issues,
-
ige. Originally debuted at Carnegie Hall on January 23rd, 1943, the composition received mixed, if not critical feedback.⏎Much of that criticism weighed on the "pretentiousness" such a narrative composition implies. As
-
mechanical engineering, better enable its practitioners to contribute to the well-being of humankind.⏎ASME Core Values⏎In performing its mission, ASME adheres to these core values:⏎•Embrace integrity and ethical conduct⏎•Embrace diversity and
-
natural gas. Higher pressures result in coal and oil.⏎Leaking from natural gas pipelines contributes to some of the atmospheric green house gas contribution.⏎Other contribution arises from that organic breakdown that occurs from the natural fermentation process in landfills. Methane<|endoftext|>
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
- L6:1135 suppresses
- L6:1218 suppresses
- L4:468 excites "that" after "than" or "like"
- L6:1458 suppresses
- L7:2537 weakly suppresses
- L5:2517 weakly suppresses
- L7:2482 weakly suppresses
- L4:2798 weakly excites
- L1:1456 weakly excites
- L5:2733 weakly suppresses
- L7:460 weakly suppresses
- L7:236 weakly suppresses
- L5:2376 weakly suppresses
- L2:974 weakly excites
- L3:2753 weakly suppresses
- L1:1084 weakly excites period in U.S. and initials
- L7:2650 weakly excites
- L4:270 weakly suppresses
- L5:610 weakly excites
- L0:2775 weakly suppresses
- L4:1733 weakly suppresses
- L6:1093 weakly suppresses
- L6:1999 weakly excites
- L2:2359 weakly suppresses
Feeds later layers
- L9:864 excites
- L9:175 excites
- L16:2660 excites
- L10:1103 excites
- L16:2861 excites
- L10:1389 excites
- L17:221 excites
- L11:1318 excites
- L12:1761 weakly excites
- L16:2466 weakly excites
- L15:2242 weakly excites
- L13:2601 weakly excites
- L9:82 weakly excites
- L12:1805 weakly suppresses
- L23:1119 weakly suppresses
- L11:2741 weakly excites
- L14:1956 weakly excites
- L10:376 weakly suppresses
- L17:1200 weakly suppresses
- L12:2962 weakly suppresses
- L9:1215 weakly excites
- L17:2011 weakly excites
- L21:1484 weakly suppresses comma after a sentence subject
- L14:2442 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
- íc
- 作出了
- .MoveNext
- 作出
- 其他玩家
- RenderWindow
- 的心理
- uation
- 作了
- uing
Suppresses
- uples
- resto
- ndef
- 맨
- System
- acie
- imitives
- тки
- _System
- ropy