unclear: electrons and code references
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. 7 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 a mixed set of words, mostly 'electron' and 'pattern matching', plus 'reference' and 'local variable'. It is not obvious what links them. Turning it up gave answers that wandered without a clear theme.
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.
-
temperatures are the factors of electrophilic substitution on aromatic compounds.⏎The most common reaction of aromatic compounds is electrophilic substitution. It is when an electron-poor reagent (an electrophile, E+) reacts with the electron
-
, respectively), but⏎Bool doesn’t have such a function. All⏎Bool is stuck with is the build-in⏎if then else-syntax, or worse: pattern matching.⏎With a ‘fold’ function (also known as
-
. It is the number of protons in the nucleus minus the number of electrons in between the nucleus and the electron in question. Basically, the nucleus attracts an electron, but other electrons in lower shells repel it (opposites attract, likes
-
shielding effect is. Essentially, the core electrons shield the valence electrons from the positive charge of the nucleus.⏎Electron-Electron Repulsions ⏎When two electrons are in the same shell, they will repel each other slightly. This effect
-
believe that pattern matching has been proved to be a useful and expressive tool in various modern languages. In particular, many aspects of this PEP were inspired by how pattern matching works in Rust and Scala .⏎Both those languages are statically typed, which
-
setting. This repair method⏎The above information equips you with everything you need to know about FCAW as a new welder. The techniques and advantages and disadvantages of FCA
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 of the elderly" (Problem of the Elderly) with regard to the elderly and the management staff. The Problem of the Elderly is a common problem. Please, the management staff, which is for the people. I
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
residence.⏎Having a professionally installed, monitored fire alarm system in your home offers you the best overall protection. When a fire is sensed, your alarm will activate all evacuation horns throughout your home. Even if you are on another level of your home,
-
, provide a source of natural ventilation, and can become a lifesaver in an emergency. If there is a fire in your home, for example, a single stairwell may become unusable, leaving family members and tenants without a clear path to
-
- The boy, that was playing in the street, that got hit by a car was my brother.⏎- My neighbor, that lives on the 2nd floor, is going to take me to school tomorrow.<|endoftext|>In the early 1
-
won’t bite you, but this still doesn’t reduce your anxiety. Likewise, you may know that it is safe to be out on a balcony in a high-rise block, yet, feeling terrified to go out on it or even enjoy the view
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
- L4:1812 suppresses
- L6:1144 excites
- L6:681 excites
- L6:2557 weakly suppresses
- L2:1657 weakly excites
- L3:1239 weakly suppresses
- L1:2948 weakly excites
- L6:776 weakly suppresses
- L6:983 weakly excites
- L4:2708 weakly suppresses
- L6:1432 weakly excites
- L6:1659 weakly excites
- L6:953 weakly suppresses
- L6:531 weakly suppresses
- L6:2868 weakly excites
- L6:2584 weakly suppresses
- L2:366 weakly suppresses
- L6:744 weakly excites
- L4:903 weakly suppresses
- L6:2110 weakly suppresses
- L6:1733 weakly suppresses
- L6:1355 weakly excites
- L5:1950 weakly suppresses unclear: 'Lorenz' and name fragments
- L6:1420 weakly suppresses
Feeds later layers
- L10:1563 suppresses
- L20:690 excites
- L8:2762 weakly excites
- L10:2572 weakly excites
- L21:2226 weakly excites
- L22:2553 weakly excites
- L25:1190 weakly excites
- L17:687 weakly excites
- L24:3049 weakly excites
- L9:1814 weakly suppresses
- L9:711 weakly excites
- L9:2832 weakly suppresses
- L25:2315 weakly suppresses
- L25:1065 weakly suppresses
- L10:1864 weakly excites
- L8:1687 weakly suppresses
- L21:1891 weakly excites
- L8:2549 weakly excites
- L26:746 weakly suppresses unclear: mixed word-start fragments
- L25:975 weakly excites
- L15:1133 weakly excites
- L20:876 weakly suppresses
- L17:2837 weakly excites
- L9:1392 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
- рон
- 去了
- -binding
- ythe
- ariat
- zar
- ��
- zek
- inston
- aleza
Suppresses
- 读后
- ulates
- ,),⏎
- walls
- slime
- .getContent
- asters
- 涞
- duct
- 一期