unclear: mixed word-start fragments
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. 26 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
It fires on a scattering of word beginnings and sounds, such as the 'N' in 'Nausea' or 'ren' in 'renounced', with no clear shared theme. Turning it down made the model get stuck repeating 'light', and turning it up gave ordinary text, so its role is hard to pin down.
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.
-
agus and out of the mouth.⏎Emesis; Vomiting; Stomach upset; Upset stomach⏎Many common problems may cause nausea and vomiting:⏎Nausea and vomiting may also be early warning signs
-
security on farms⏎Biosecurity is about reducing the risk of disease transmission. We know that bTB can be transmitted by direct contact (e.g. nose-to-nose) or indirect contact (e.g. via contaminated pasture) but simple exclusion
-
- Sweating, clammy skin⏎- Enlarged (dilated) pupils⏎- Insomnia (sleeping difficulty)⏎- Loss of appetite⏎- Nausea and vomiting
-
treated thus he renounced all pardon, as he had never thought to have offended, and only asked
-
abdomen, including the stomach. This can lead to a feeling of fullness, even if you haven't eaten much, and can make you lose your appetite.⏎Nausea is a feeling of un
-
they mistake bright-colored pieces for food. Sea turtles and migrating birds can become entangled in abandoned plastic fishing nets known as “ghost nets.” Plastics also can leach harmful chemicals into the water which can be carried along the food chain back 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: light refraightment refightment refightment refightment refightment refightment refightment refightment refightment refightment refightment refightment refightment refightment refightment ref
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
based on the borrower's income, expenses, and credit history. The results of a stress test can impact the amount a borrower is able to borrow for a mortgage and the interest rate they receive.⏎Debt Servicing Ratio
-
how much interest to charge. By not knowing their credit score, people risk paying more interest when they take out a loan and may face limits on the amount they can borrow. In extreme circumstances, they may find themselves “locked out” of credit
-
also known as the “Poop to Power” project. The station will provide fuel for the first natural gas vehicles in the City’s fleet. Initially, vehicles will fill up on natural gas from conventional sources, but once the full project is complete
-
nett Town site Company dedicated the streets, avenues and alleys to the use of the public. The right and privilege to construct gas, water and sewer lines, to erect telephone and electric light poles was reserved to the company. A small town site
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
- L23:53 excites
- L23:373 suppresses 'to' after a noun or adjective
- L22:627 suppresses
- L19:2632 suppresses
- L23:491 excites
- L17:645 excites
- L11:223 suppresses
- L17:1361 suppresses 'that' after verbs like recommend
- L4:113 suppresses
- L6:1133 excites "to" before an infinitive verb
- L21:536 excites
- L16:1153 weakly suppresses
- L7:520 weakly excites
- L18:178 weakly suppresses
- L20:349 weakly excites
- L19:1830 weakly excites
- L18:971 weakly suppresses
- L9:226 weakly suppresses 'to' starting an infinitive
- L15:184 weakly suppresses
- L7:681 weakly excites 'from' after prevent or stop
- L21:1663 weakly excites
- L21:1657 weakly suppresses
- L0:1018 weakly excites
- L6:1880 weakly excites "by" introducing a deadline or means
Feeds later layers
- L27:2060 weakly suppresses
- L27:3028 weakly suppresses
- L27:1972 weakly excites
- L27:1328 weakly suppresses
- L27:2991 weakly excites
- L27:2653 weakly suppresses
- L27:1437 weakly suppresses
- L27:1414 weakly excites
- L27:2082 weakly excites
- L27:1048 weakly suppresses
- L27:1705 weakly excites
- L27:973 weakly suppresses
- L27:1041 weakly excites
- L27:3008 weakly excites
- L27:1231 weakly suppresses
- L27:1471 weakly suppresses
- L27:330 weakly excites
- L27:2322 weakly excites
- L27:560 weakly excites
- L27:1341 weakly excites
- L27:2519 weakly excites
- L27:1762 weakly excites
- L27:2816 weakly excites
- L27:2111 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
- 地上
- 思う
- perish
- |_|
- xfc
- 地说
- xFFFF
- 途
- Err
- 床上
Suppresses
- чь
- .depth
- ight
- edReader
- ző
- 蜇
- .char
- bbing
- bay
- rite