'on' after pressure or strain 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. 8 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 'on' when it comes right after phrases like 'put strain on', 'place pressure on' or 'put demand on', where something is being loaded or stressed. When it was turned up, the model's writing drifted toward talk about pressure and stress on things, though the steered text was often repetitive.
Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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.
-
DISCS⏎The entire spinal cord is sensitive due to the surrounding nerves. A herniated disc usually causes chronic back pain because of the extra pressure and stress put on the back. Patients ultimately need to rest the body and the spine;<|endoftext|>F
-
reintroduced the<|endoftext|>Jogging is a well known activity that can have hazardous effects on the knee while running on pavement. A tremendous amount of strain is placed on the knee when running on a hard surface, like asphalt or pavement. While city
-
of the database rather than re-generating it from scratch. The result is that hundreds of queries are replaced with one single query, thereby significantly lightening the load on the web server.⏎I assume it’s
-
, and weaker scar tissue sometimes replaces muscular tissue in these cases. Obesity, lack of use, heavy lifting, chronic cough, and chronic constipation place additional pressures on the pelvic floor. Genetically weak collagen, and smoking (which weakens collagen
-
and contributing factor to plantar fasciitis. Whether you are gaining body fat, muscle mass, or healthy weight from pregnancy, the added pounds put extra strain on your feet. This can cause plantar fasciitis for the first time,
-
problem so that the spark plugs can ignite the proper mixture.⏎3) Engine Drags⏎Whenever you increase the load of your vehicle, it will put more demand on the engine to
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 a new idea about the use of a special type of coin. The price is higher. I have a special type of coin. I have a new idea. The price is higher. The price is higher. The price is higher
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
to be used as a concrete set accelerator⏎- No flocculation or coagulation⏎- All concretes and mortar containing steel reinforcement, post-tensioned concrete, concrete in contact with chlorides, parking structures, piers,
-
the source code and object file is located at sub folder similar to your project name, the executable file is under sub folder debug.⏎Using Borland C++ Compiler Command Line Tools⏎- Installing and running the
-
C = Classes (swim lessons and CPR)⏎- D = Devices (personal flotation devices (PFD’s), life jackets, and rescue tools)⏎A = Adult Supervision⏎Drowning is a leading cause of unintentional death in Sacramento County
-
launched to the International Space Station (ISS). Samples from Prof. Christian Betzel’s research group at Universität Hamburg will be aboard to be analyzed in the absence of gravity.⏎The experiment in the context of the CUI research is expected to be conducted
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
- L6:2663 excites
- L4:2351 excites
- L7:2901 excites
- L5:1683 excites
- L7:2849 weakly excites
- L7:2435 weakly suppresses
- L6:2496 weakly suppresses
- L4:1235 weakly excites
- L6:2986 weakly suppresses
- L5:2745 weakly suppresses
- L6:946 weakly excites
- L7:2218 weakly suppresses
- L6:3066 weakly suppresses
- L7:1877 weakly excites
- L6:1652 weakly suppresses
- L3:1064 weakly suppresses
- L7:1538 weakly excites
- L7:2612 weakly excites
- L4:1753 weakly suppresses
- L6:2154 weakly excites
- L4:2439 weakly suppresses
- L5:1494 weakly suppresses
- L3:1185 weakly excites
- L0:311 weakly suppresses
Feeds later layers
- L10:1079 excites
- L11:993 excites
- L23:2958 excites
- L9:664 suppresses
- L10:2602 excites
- L23:1750 excites
- L10:2024 excites
- L12:2994 excites
- L10:274 weakly suppresses
- L19:2877 weakly excites
- L9:1473 weakly excites
- L20:3012 weakly excites
- L10:2919 weakly excites
- L20:2641 weakly excites
- L9:1807 weakly excites unclear: mixed, health lists and 'weight'
- L23:1954 weakly suppresses
- L11:2206 weakly suppresses
- L12:2169 weakly excites
- L11:749 weakly excites
- L11:1737 weakly suppresses
- L15:2011 weakly excites
- L20:2246 weakly excites
- L23:2345 weakly excites
- L10:1207 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
- lict
- …)
- 减轻
- 慰
- hana
- undue
- licted
- 负荷
- eted
- ARY
Suppresses
- order
- po
- 告诉他
- .PREFERRED
- cheme
- ício
- erior
- posable
- flag
- riages