"with" after combine, mix or react verbs
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. 11 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
Fires on the word "with" when it follows a verb of joining or reacting, like "combined with", "mixed with" or "reacts with". Turning it up made the model drift into rambling, unrelated text.
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.
-
through a wire to create an electric current. The electrons complete the detour to meet up with the ions in the final segment of the fuel cell where they combine with oxygen to create water or carbon dioxide. The energy produced is clean, with water
-
BODY.⏎Now upon death, the intelligence of the person, automatically within seconds enters the body of another child that is 11 years. It then combines itself with the intelligence that child had from birth. Example, if a man became a doctor
-
is mined, melted, and split into hydrogen and oxygen by electrolysis. The lakes' hydrocarbons are pumped from the lakes into stations where they are mixed with the oxygen derived from the ice—that is to say, burned—resulting in
-
decided that the licensing provisions were constitutional and reversed the appellate court. The Jehovah’s Witnesses from Alabama appealed to the U.S. Supreme Court, which consolidated the case with two other licensing fee cases involving Witnesses.⏎Court originally upheld license requirement⏎Justice Stanley
-
there is blooming vegetation bordering the treated field. EPA is prohibiting tank mixing of sulfoxaflor with pesticides that have shown evidence of synergistic activity with sulfoxaflor. The product label directs applicators to more information and
-
“Hydrogen is quite a versatile thing,” Constantine says. “So if we’re looking at breweries in particular, they can use that hydrogen to blend with natural gas for heating or they can also convert it back into electricity for renewables if
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: "A single sentence with a single idea, and then a few sentences are distributed across the sentence structure, which is the primary unit. In the case of a single sentence, there is a single idea, and the sentence structure is typically a
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
0" long, and 8" wide⏎Attributes: Good for Containers⏎When to Sow Outside: RECOMMENDED. 4 to 6 weeks before your average last frost date, when soil temperature is at least 40°F
-
How to Plant Marigolds⏎- If you wish to start growing marigolds in spring, sow them indoors, particularly six to eight weeks before your area’s last frost date. If you are fine with later blooms, sow
-
.⏎- Use voice commands⏎- Nowadays, AI is becoming more intelligent. We can use voice commands instead of typing on the screen.⏎- Avoid sharing your devices with others⏎- If unnecessary, do not share your smartphone or other devices with others
-
014⏎One approach if giving criticism is to put it inside a sandwich- open with aspects you praise or agree with, offer critical statements, and close with a positive. Maybe the best advice is to comment in the style and mood that
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
- L6:800 excites
- L9:1252 suppresses
- L10:2213 suppresses 'to' or 'into' showing movement between two things
- L9:2267 excites 'with' after exchange nouns
- L9:1820 excites
- L8:150 excites
- L7:1949 excites
- L7:2161 excites
- L2:2260 excites 'a' after 'in' or 'such'
- L6:1828 suppresses
- L8:2999 suppresses
- L6:364 excites
- L6:1084 suppresses
- L9:878 suppresses
- L10:515 weakly suppresses
- L3:2502 weakly suppresses
- L8:1083 weakly excites "x" or "by" between two dimensions
- L5:910 weakly excites
- L9:1144 weakly excites
- L8:809 weakly excites
- L8:2596 weakly excites
- L9:1908 weakly excites
- L4:502 weakly excites
- L10:1359 weakly suppresses
Feeds later layers
- L15:2247 excites
- L14:2830 excites
- L14:2100 excites
- L14:1654 excites
- L12:2739 excites
- L19:2698 excites
- L12:258 excites
- L15:2205 excites
- L17:2769 excites
- L15:2522 excites
- L12:1394 excites
- L16:1469 excites
- L15:3021 excites
- L13:716 excites
- L13:2754 excites
- L14:2320 excites
- L12:1454 excites
- L12:2930 excites "and" joining contrasting pairs
- L19:1572 excites
- L15:1037 excites
- L18:2256 excites
- L12:1907 excites
- L12:1544 suppresses
- L15:361 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
- sake
- оборот
- Relationship
- 一起
- entifier
- работу
- enal
- 搭配
- gorit
- 同事
Suppresses
- zano
- Fra
- Chambers
- quo
- pons
- あげ
- Hawkins
- -pill
- hog
- /qu