'for' after preparation or purpose nouns
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. 1 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 "for" when it follows words like "preparation", "preparing", or "earmarked", where the next thing is a purpose or target. Its negative side is weak and shows up on "for" in loop or quantifier uses. Turning it up gave text that changed topic, so the steering doesn't clearly show its role.
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.
-
cut a large tree shape out of craft paper and affix it to the dining room wall in preparation for thanksgiving. Then, during the meal, each family member and friend turn over
-
age of eight weeks, dogs scheduled for surgery, and people with pre-existing anemia mustn't be given garlic.⏎
-
in March 2023, which brought together<|endoftext|>Before I go further, I want to emphasize that I believe that school is not just about preparing students for the future. It's about helping them manage their now. I think we often
-
15 inches of rain over a 72-hour period.⏎Readiness, Response and Recovery⏎With a forecast for heavy rainfall, the District began preparations for Tropical Storm Isaac in the days before it impacted South Florida. Water levels in regional
-
Curtiss and purchase a plane for $7,000 ($157,000 in today
-
breast—near the turkey's crop—is known as breast sponge. Gobblers develop the fatty layer of breast sponge in late winter as they feed heavily in preparation for breeding season. The fat deposit enables gobblers to have energy to spend little time
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 have been thinking about lately, and I have been thinking about lately. So I have been thinking about lately. So I have been thinking about lately. So I have been thinking about lately. So I have been thinking about lately. So
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
if n> 10⏎> 3 print "This might take a while..."⏎> 4 for i = 1 to n⏎> 5 for j = 1 to i⏎> 6 print
-
for all types of wrens. These birds can have substantial broods, often from
-
for all the battles that wound up half-won, it was a romantic time in our
-
langue was responsible⏎for the financing of the building, and it was supposed to build the auberge
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
- L0:1301 suppresses
- L0:178 suppresses
- L0:1224 suppresses
- L0:1 weakly suppresses
- L0:9 weakly suppresses
- L0:1960 weakly excites unclear: mixed medical and harsh words
- L0:3031 weakly excites
- L0:33 weakly excites
- L0:1824 weakly excites
- L0:1148 weakly suppresses
- L0:2031 weakly suppresses
- L0:1512 weakly excites
- L0:492 weakly excites
- L0:154 weakly suppresses
- L0:2538 weakly suppresses
- L0:81 weakly suppresses
- L0:97 weakly excites
- L0:1796 weakly excites
- L0:2443 weakly suppresses
- L0:1378 weakly excites
- L0:3058 weakly excites
- L0:908 weakly excites
- L0:2072 weakly excites
- L0:3 weakly suppresses
Feeds later layers
- L10:1648 weakly excites
- L14:2150 weakly suppresses
- L24:2437 weakly excites
- L11:2269 weakly suppresses
- L14:3052 weakly suppresses
- L3:2469 weakly excites
- L7:2815 weakly excites
- L15:692 weakly suppresses
- L3:2710 weakly excites
- L10:800 weakly excites
- L7:1855 weakly excites
- L7:2427 weakly suppresses word before 'type 2 diabetes' or a condition
- L26:122 weakly excites
- L11:374 weakly excites
- L11:2280 weakly suppresses
- L6:1045 weakly excites
- L7:1460 weakly suppresses
- L18:967 weakly excites
- L6:1660 weakly suppresses
- L24:289 weakly excites
- L13:822 weakly suppresses
- L8:2918 weakly suppresses
- L18:552 weakly excites
- L12:1474 weakly suppresses
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
- 什么样的
- -side
- etched
- -sided
- médico
- ей
- ós
- :\/\/
- pared
- (),"
Suppresses
- commod
- Ups
- Institutes
- ButtonItem
- я
- ji
- Channels
- issa
- CHANNEL
- nty