unclear: words before 'and' or 'or'
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. 18 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
This neuron fires on words that sit in the middle of a list or contrast, often just before 'and' or 'or', as in 'multidisciplinary engineering' or 'exercise or health program'. The pattern is loose. Turning it up made the writing repeat itself and drift.
Written by Claude from this neuron's activations and steering outputs. Confidence: low. 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 American Society of Mechanical Engineers, ASME is a not-for-profit professional organization promoting the art, science and practice of mechanical and multidisciplinary engineering and allied sciences. ASME develops codes and standards that enhance public safety,
-
|Math Topics:||Graphing of Data|⏎© 1994- Drexel University. All rights reserved.⏎The Math Forum is a research and educational enterprise of the Drexel University School of Education.<|endoftext|>Maybe you’ve
-
(more or less). Thus began the era of dancing babies and other horrors best forgotten<|endoftext|>PROTEOMICS is the study of many proteins at once- or a description of proteins in a cell, tissue or cellular organelle on a global
-
of the theory, linking it to best practices in the classroom.??<|endoftext|>I conducted this project as part of my doctoral dissertation in 2018 and early 2019 in an art immersion school in Western Canada.⏎The goal
-
not intended to substitute for the diagnosis, treatment, or advice of a medical professional. Consult your medical professional for guidance before changing or undertaking any diet, exercise or health program, or before using any prescription or over-the-counter drug. Neither the author
-
. Sellers and buyers can share hundreds and thousands of kilometers. Any distance is always costs, logistics. And it is very difficult to understand exactly what exactly and in what volumes buyers of another region want
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 problem is a simple one, but the first question is a simple question that asked it at first, which was originally a simple question, but it was a question that was initially a simple question. For example, I was once asked a
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
morality of God, and see him as their leader and prime spiritual guide.⏎Although the overwhelming majority of Sufis, both pre-modern and modern, were and are adherents
-
actionable insights. By understanding the correlation between CPPs and product’s critical quality attributes (CQAs) – measured in real time in-line, on-line and /or at-line by suitable analytical instruments, the process can be controlled in real-time
-
Sensitive Person.⏎- This may mean that you fall on the right side of the bell curve as a gifted person.⏎- This may mean you do have or are mislabeled as having a mental illness such as Emotion Regulation Disorder, ADHD
-
they are, and that the issues they bring to therapy are not restricted to (or located) within the clients themselves, but rather are⏎What are the key concepts of
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
- L16:2994 strongly suppresses
- L10:163 suppresses
- L10:74 excites
- L17:1575 excites
- L16:678 suppresses
- L15:1443 excites
- L16:1388 suppresses
- L15:784 excites
- L13:1554 excites
- L15:217 suppresses
- L17:2221 excites
- L14:745 suppresses
- L9:92 suppresses "and" or "or" joining two items
- L13:131 excites 'and' joining parallel phrases
- L4:57 suppresses
- L17:1537 suppresses
- L9:212 excites
- L10:1643 suppresses
- L14:1249 suppresses
- L16:1937 suppresses 'and' joining two time or quantity ranges
- L13:372 suppresses
- L16:1659 excites
- L16:582 excites
- L15:1459 excites
Feeds later layers
- L21:735 excites 'or' in ranges and alternatives
- L19:883 excites
- L19:1431 excites
- L19:1849 weakly excites
- L19:2711 weakly excites
- L25:104 weakly excites
- L19:1823 weakly excites 'and' inside organisation or paired names
- L19:2311 weakly excites
- L19:2562 weakly excites
- L22:1113 weakly suppresses
- L19:1675 weakly excites
- L23:2414 weakly excites
- L25:1694 weakly excites 'and' joining paired names or fields
- L21:1402 weakly excites
- L20:867 weakly excites
- L22:2115 weakly suppresses
- L26:124 weakly suppresses
- L21:1540 weakly excites
- L27:49 weakly suppresses
- L21:510 weakly suppresses
- L20:1458 weakly excites
- L20:2027 weakly excites
- L19:1673 weakly excites
- L23:2356 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
- rew
- 蔼
- éro
- 伺
- люч
- 잉
- fé
- ış
- loser
- 其他的
Suppresses
- is
- iani
- eties
- etri
- ');?>"
- ։
- ationale
- Official
- internal
- angel