acronyms and 'case' in running text
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. 3 of Qwen3-0.6B. Its strongest snippets are fairly consistent: several of them 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 responds to a few scattered words: acronyms like 'OAS' and 'MDC', and the word 'case' in phrases like 'the case'. The examples do not share a clear theme, so it is hard to say what unifies them.
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.
-
0 students, and has been expanded to include Argentina and Brazil with the support of the Organisation of American States (OAS).⏎The Organic Gardens programme recommends the use of greenhouse-style protected gardens so that children
-
a dog’s leg and break the ankle of a human, and the poison causes gangrene if untreated. They live along the banks of rivers and oasises.⏎They ambush creatures, hiding in the sand and snapping when a foot or
-
debt to women in dictatorships.”⏎Now, the CIDH has brought the ESMA exhibit to the Organization of American States (OAS) Headquarters in Washington DC in an attempt to start a conversation about crimes against women
-
and pancreatic cancer.⏎A family history of breast cancer, male or female, increases the man’s chance of inheriting the disease, though this isn’t always the case — but as with women, many men have no obvious
-
they just talk past each other. An additional challenge for a mediator, on top of getting the parties to listen to one another, is that it’s often the case that a person
-
have made sure that they left the mark on the city be designing different pieces of beautiful architecture that are itself an amazing piece of work.⏎Delhi will drive anyone crazy when get a views of its age old style and zealous past. It is
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: that is, a problem of their own. So, the problem is to find out how much time the problem has taken. So, the problem is to find out how much time the problem has taken. So, the problem is to find
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
INO - ARMOUR - ARTWORK - BIOLOGY - BLACK BOX - COMPUTERS - ELECTRONICS - ENERGY - FRAME - HEAD - JAWS - JIMMY WATSON - KITS - LEGS - MECHANICS - MOTORS
-
?⏎- What are high fidelity prototypes?⏎- What are the three key points in a relationship framework?⏎- Why is it called LOFI?⏎- What does Lo Fi mean?⏎- What is a wireframe diagram?⏎- When should I use low
-
, and more than one story is recorded of people succumbing to hunger and disease while still in the harbor.⏎The remarkable confluence of four men made the Jeanie Johnston a exception to this rule. In the New World, there was John M
-
writing of the user⏎instructions. Completed CAD models are also used in the wider⏎organization, such as for marketing and packaging.⏎source : wikipedia⏎How lego used to do it⏎do you know?⏎How people do it themselves⏎L
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
- L1:2336 weakly excites
- L2:1321 weakly suppresses
- L0:2814 weakly excites
- L2:1012 weakly excites
- L1:1574 weakly suppresses
- L2:1064 weakly suppresses
- L0:1038 weakly excites
- L0:741 weakly suppresses
- L1:128 weakly excites
- L2:111 weakly suppresses
- L0:797 weakly suppresses
- L1:1154 weakly excites
- L0:1301 weakly suppresses
- L2:2790 weakly suppresses
- L1:491 weakly excites
- L1:2358 weakly suppresses
- L2:2271 weakly suppresses 'in' in 'take in' phrases
- L0:552 weakly excites
- L1:2643 weakly suppresses
- L2:1494 weakly suppresses
- L2:1790 weakly excites
- L1:1327 weakly suppresses
- L1:571 weakly excites
- L1:2176 weakly excites
Feeds later layers
- L4:146 excites
- L8:2783 weakly excites
- L5:2799 weakly excites
- L23:1227 weakly suppresses
- L13:473 weakly excites
- L17:2555 weakly excites
- L20:263 weakly suppresses
- L13:1987 weakly suppresses
- L19:2337 weakly excites
- L9:1187 weakly excites
- L4:991 weakly suppresses
- L12:2126 weakly suppresses
- L5:90 weakly excites
- L5:2456 weakly excites
- L21:1358 weakly suppresses
- L12:1338 weakly suppresses
- L15:2189 weakly suppresses
- L10:2857 weakly suppresses
- L9:2024 weakly excites
- L27:2990 weakly suppresses
- L6:2612 weakly suppresses
- L18:2937 weakly excites
- L19:2303 weakly suppresses
- L19:223 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
- ận
- afe
- active
- embre
- isse
- lder
- eed
- צו
- lide
- olar
Suppresses
- ;',
- inds
- isposable
- キ
- strings
- (savedInstanceState
- 膑
- InstanceState
- ail
- caul