"from" after hide, shield, or safe
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 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
The neuron fires on the word "from" when it follows words about hiding or protection, such as "hide from", "hidden from view", "safe from" or "shield from". Turning it up did not produce text about hiding, so the steering does not confirm the idea.
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.
-
is that its vast forested areas gave them ample areas to hide from man (until recent centuries).⏎Why is Africa known for its animals?⏎Africa is
-
several hours, they reach the meeting place at the head of the bay behind several small islands. It was a nice sheltered spot. It would even be safe from the occasional sudden great storms that struck without warning in that late summer season. As
-
the early 20th century, the mountains and hollers of Polk County were infamous for hosting moonshiners and bootleggers, hidden away from the long arm of the government “revenooers.” But for the past
-
This is, actually, just the most popular use of files.⏎Apart from text files, we have a range of system files whose content is hidden from the user's eye but is crucial for the proper functioning of websites and applications.
-
the victim only. A bully’s threats, lies, and put-downs can be posted online and passed along to everyone and anyone. There is no safe haven from a bully.⏎- Cyberbullying is more difficult to stop. When a bully
-
deposits.⏎Hydrophobin assemblies shield the human pathogenic fungus Aspergillus fumigatus from the immune response during infection, allowing this fungus to go on to cause life-threatening
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: https://www.abc.com/222/1000/1000/1000/1000/1000/1000/1000/10
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
contain that have been useful in research.⏎The glands are enclosed in a capsule of and internally divided into . Blood vessels and nerves enter the glands at the and gradually branch out into the lobules.⏎In the duct system, the lumens formed by
-
<|endoftext|>A metal pole anchored down with cables juts out of the sand next to a public walkway at First Landing State Park. Four small rods stick out horizontally from the sides.⏎Research scientists from Old Dominion University installed the contraption on the d
-
movements, the neurobiologists concentrated on a group of around 20 neurons which send out their extensions from the midbrain to the spinal cord. The scientists already knew that the cells
-
lifting.⏎The spinal cord runs through the spinal canal, a tunnel that is formed by the holes in the center of each of the vertebrae. Nerves branch from the spinal cord, pass through openings (foraminae) between the vert
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
- L4:192 excites
- L6:2657 suppresses
- L6:2836 weakly excites
- L1:1676 weakly suppresses
- L6:2743 weakly excites
- L4:2269 weakly suppresses
- L6:1512 weakly excites
- L6:1192 weakly excites
- L5:28 weakly excites
- L6:545 weakly excites
- L6:2595 weakly suppresses
- L6:1297 weakly excites
- L3:2791 weakly excites
- L2:1237 weakly excites
- L6:1442 weakly excites
- L4:3019 weakly suppresses
- L5:141 weakly excites
- L5:2520 weakly suppresses
- L2:1180 weakly excites
- L4:1469 weakly suppresses
- L2:2108 weakly excites
- L6:1799 weakly excites
- L0:1290 weakly suppresses
- L1:2488 weakly suppresses
Feeds later layers
- L8:2797 excites
- L8:2879 suppresses
- L12:1276 suppresses
- L8:427 weakly suppresses
- L10:852 weakly suppresses
- L11:2369 weakly suppresses
- L10:2452 weakly suppresses
- L20:1994 weakly suppresses
- L23:1822 weakly suppresses
- L10:3012 weakly suppresses
- L11:337 weakly excites
- L8:918 weakly suppresses
- L9:1619 weakly suppresses
- L11:2214 weakly suppresses
- L11:1716 weakly suppresses
- L17:1375 weakly suppresses
- L22:2193 weakly suppresses
- L23:125 weakly suppresses
- L11:2047 weakly suppresses
- L17:2325 weakly excites
- L11:422 weakly excites
- L8:2798 weakly suppresses
- L19:1931 weakly suppresses
- L20:1491 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
- ummies
- UNUSED
- IData
- .GetData
- .READ
- 有色
- avoidance
- otas
- avadoc
- alloca
Suppresses
- adder
- reet
- arse
- UES
- aight
- uter
- anks
- erra
- idency
- 沿着