'or' linking two alternatives
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. 16 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 'or' when it joins two alternative names for the same thing, such as 'boundary or interface', 'patella or the kneecap' or 'symbol or underlying'. It is most common in technical explanations. Turning it up or down did not produce a clear, related 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.
-
single ray of light strikes a boundary between two mediums and is refracted.⏎- The light ray which travels through the incident medium or first medium and strikes the boundary or interface is called the incident ray.⏎- The light ray which travels into the second
-
the vastus lateralis, vastus medialis, vastus intermedius, and the rectus femoris. These muscles are present just above the patella or the kneecap and form the patellar tendon. These muscles perform the function
-
options contract has the following properties or attributes:⏎- Expiration Date⏎- Symbol or Underlying⏎- Strike Price⏎- Type: Call or Put⏎The contract
-
<|endoftext|>Cancellous bone, also called trabecular bone or spongy bone, light, porous bone enclosing numerous large spaces that give a honeycombed or spongy appearance. The bone matrix, or framework, is organized into a
-
one is redundancy, so if one drive breaks down, the data will be accessible from the remaining ones, and the second is improved performance as the input/output, or reading/writing operations will be distributed among different drives. There are different RAID types
-
which represents the final prediction or output.⏎C. Splitting Data based on Features⏎The key feature of decision trees is their ability to split data based on features or
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: I want to be able to have a small part of a small part of a small part of a small part of a small part of a small part of a small part of a small part of a small part of a small part of a
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
The first self-evident truth of the Declaration of Independence says that “all men are created equal.” No words in the American lexicon are as eloquent or as powerful. We believe this to be a fundamental moral truth. Our Constitution begins
-
standard Teletype paper tape uses only 8 bits.) Cities would hire private companies to translate the data into reports that traffic engineers could use to adjust traffic lights or improve roads.⏎Bill Gates and Paul Allen were high school students at Lakeside School
-
short burst of time,, it can be overwhelming. This is because neurodivergent brains process information differently, making it hard for them to switch between tasks quickly or efficiently in stressful situations.⏎Neurodivergent individuals may also become distracted or
-
⏎Machine Learning can be performed using multiple approaches. Even if you are not planning to become a machine learning developer, knowing these techniques can enable you how your current or future AI project will be built.⏎The four basic techniques of ML are: supervised
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
- L15:3059 suppresses 'and' or 'or' joining the last list items
- L12:1443 excites opening parenthesis after a name or term
- L6:990 excites
- L10:135 excites 'or' in alternative requirements
- L15:727 excites
- L15:2935 suppresses
- L9:2010 suppresses
- L12:2076 excites
- L9:92 suppresses "and" or "or" joining two items
- L14:50 excites
- L10:144 suppresses
- L14:2276 excites opening bracket after a term or number
- L3:474 excites
- L15:515 suppresses word before 'that', 'to' or 'for' continuation
- L5:330 excites opening parenthesis after a noun
- L15:1907 excites the word 'or' between alternatives
- L10:1459 suppresses
- L9:1065 weakly suppresses
- L14:2603 weakly excites
- L15:693 weakly excites commas inside numbers and asides
- L12:2252 weakly suppresses
- L14:1209 weakly excites
- L11:406 weakly suppresses
- L10:1643 weakly suppresses
Feeds later layers
- L17:2274 excites
- L18:2544 excites
- L17:2704 excites
- L17:2258 excites
- L17:3025 excites
- L18:1916 excites
- L18:192 suppresses
- L19:2867 excites unclear: 'or something' and open parentheses
- L19:331 suppresses
- L18:933 excites
- L19:451 suppresses
- L19:1675 weakly excites
- L19:356 weakly suppresses
- L17:1214 weakly excites
- L21:1402 weakly excites
- L19:1438 weakly excites
- L20:1339 weakly excites
- L19:354 weakly suppresses
- L18:2136 weakly excites
- L17:1075 weakly suppresses
- L19:2711 weakly excites
- L17:667 weakly excites
- L20:2794 weakly excites
- L17:180 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
- itch
- ��
- yah
- hay
- phans
- công
- الماضي
- scratch
- prech
- insic
Suppresses
- ameleon
- 門
- 美的
- AdapterManager
- }}>⏎
- 门
- *>(
- ][/
- .SuspendLayout
- 花开