'more' or 'not' before a likelihood claim
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. 20 of Qwen3-0.6B. Its strongest snippets are fairly consistent: several of them 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 words like 'more', 'not' and 'likely' when the sentence is about one thing being more or less likely to happen, such as being more prone to injury. Turned up, the model drifted into repeating 'will' and 'going to', which fits its future-tense pull.
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.
-
the nerve input and how long it is lasting for a particular contraction. It is used to assess whether or not the muscle is spastic, since weak muscles are more prone to injury and spasm. Bulbocavernosus reflex latency is
-
’s mental state. Cognitive empathy is the ability to understand that mental state. These components are independent of each other, which means a person capable of emotional empathy isn’t necessarily capable of cognitive empathy, and vice versa.⏎What Are the Issues Concerning
-
scary, but it’s not. If you’re off a few (or 30) seconds or your water’s a couple degrees off temperature, you’re NOT going to ruin your film. The more accurate you are,
-
metabolism slows down, allowing it to conserve oxygen and stay underwater longer. Similarly, the activity level of the turtle can affect its breath-holding capabilities. A resting turtle can hold its breath for longer than an active one.⏎Do turtles need to breathe
-
at tasks that they are fully able to handle successfully.⏎* Those individuals who have developed learned helplessness are far more likely to suffer from depression.⏎* Such people will tend to feel unsatisfied in
-
⏎There are useful variants too (shown here without redirection)⏎Note: even if svn can make the difference of another directory (svn diff mydirectory), it is not recommended to do<|endoftext|>by Clinton Wee⏎Being one of the only atheist countries
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 are some 2D that are not the best in the world, but they have the same or better features as the best 1D. What is the best 2D that we have that we are supposed to have that we
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
<|endoftext|>The nail matrix is the layer of keratin-producing cells at the base of each toe or finger. As new cells arise, the hard husks of older cells are pushed forward, creating the nail plate – what we think of as just the
-
s and apply it to the male barnyard fowl. Slang doesn’t usually have a euphemistic quality. When we think of euphemisms, we think of words that are substituted because their connotations are less distressing than the
-
transmission of DNA via mobile genetic element is quite different from the well-known means by which parents pass on their genetic material to their offspring. In animals, pieces of DNA typically cannot jump from one adult organism's genetic material to another adult organism's genetic
-
means looking not only at political options but also at the way we humans are set up — how we’re wired. When public policies don’t account for that, we have reduced horizons, diminished resources, and polarization.⏎How are we wired?
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
- L19:2133 strongly excites
- L17:1118 suppresses mid-clause words in technical prose
- L19:1862 suppresses
- L14:2571 excites
- L17:730 excites
- L19:553 suppresses
- L15:1829 excites
- L19:545 suppresses
- L19:2157 suppresses
- L15:1113 suppresses
- L19:564 suppresses
- L12:1702 excites
- L17:1356 excites
- L14:1347 weakly suppresses
- L8:2534 weakly suppresses
- L16:1223 weakly excites
- L18:233 weakly excites
- L10:197 weakly suppresses
- L6:1436 weakly suppresses "from" after a change in amount
- L9:281 weakly suppresses
- L2:270 weakly suppresses
- L12:385 weakly excites
- L19:2475 weakly suppresses
- L15:1513 weakly suppresses
Feeds later layers
- L23:765 strongly excites
- L27:1015 strongly excites
- L25:2368 strongly excites
- L23:1555 strongly excites
- L22:745 strongly suppresses
- L23:1173 excites
- L26:2680 excites
- L24:1106 excites
- L22:2824 excites
- L23:103 suppresses
- L21:1699 excites
- L22:630 excites
- L21:2523 excites
- L22:2166 excites
- L25:2151 excites
- L21:2121 excites
- L23:2305 excites
- L23:1419 excites
- L24:277 suppresses
- L22:2224 weakly excites
- L22:2399 weakly excites
- L26:1844 weakly suppresses
- L26:2147 weakly suppresses
- L22:1616 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
- will
- will
- 将会
- Will
- ’ll
- Will
- 将在
- 'll
- 将
- WILL
Suppresses
- reon
- ают
- ])]⏎
- 炳
- ennent
- ]){⏎
- ائم
- jq
- doesn
- ious