'as' in 'as ... as possible'
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. 23 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 'as' in phrases like 'as quickly as possible' or 'as simple as possible'. When it was turned down, the model got stuck repeating the word 'possible' over and over.
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.
-
member companies, other industries and with federal, state and local governments to prepare for hurricanes and return operations as quickly and as safely as possible.⏎API member companies also independently work to improve preparedness for hurricanes and other
-
be certified by a physician, licensed clinical psychologist or psychiatrist.⏎The goal of homebound instruction is to facilitate the student’s return to the current classroom setting as soon as medically indicated. Homebound instruction is temporary. Requests for homebound instruction extending beyond
-
, or surgery that will cure osteogenesis imperfecta (OI). The goal of treatment is to prevent deformities and fractures and allow the child to function as independently as possible. Treatments for preventing or correcting symptoms may include the following:⏎Management of
-
base Alpha was developed in support of those goals."⏎In addition to funding the game, NASA also had a hand in ensuring that it was as accurate as possible, going so far as to have several engineers and scientists review the game for
-
"deferred" items in relation to accounting transactions, we begin to go further into the reporting side of accounting, and in order to keep this article as simple as possible, we are not going to expand on the relationship of deferred items to balance
-
cattle have shown it is as effective as current vaccines. Whilst a commercial product is still several years away, the team hopes that the technology can be transferred as quickly as possible to make it available to a global market.⏎One of the problems of existing
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: what is the number of distinct self-possible-possible-POPOPOPOPOPOPOPOPOPOPOPOPOPOPOPOPOPOPOPOPOPOPOPOPOPOPOPOPOPOPOPOPOPOPOPO
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
able to participate in the planning and execution of treatment plans. The Merchant of Venice is currently undergoing the prescribed treatment plan.⏎The conservation of these historical works has been made possible through the generosity of the Scripps
-
advent of deep learning is changing this landscape and putting automation in reach. Resources such as open-source libraries, Nvidia hardware, and FLIR cameras are helping to make this change happen.⏎Deep Learning Made Easier⏎Deep learning is a powerful tool for
-
incorrectly, an oil-analysis program is a waste of time and money. Key guidelines include:⏎- Take warmed-up machine samples while equipment is running. If that’s not possible, sample no later than 30 minutes after shutdown⏎- Use clean
-
taking steps to ensure that water and air quality are adequately protected during surface and subsurface natural gas development activities.⏎The recent increases in domestic natural gas supplies have been made possible by two technologies—horizontal drilling and hydraulic fracturing—that allow energy companies to
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
- L18:2565 excites
- L20:2994 excites
- L21:2006 excites "as ... as possible" phrasing
- L19:1307 excites
- L22:550 suppresses
- L19:2765 excites
- L22:1104 suppresses
- L20:1501 weakly excites
- L13:2522 weakly excites
- L15:1930 weakly suppresses
- L19:292 weakly excites
- L7:63 weakly excites
- L18:3049 weakly suppresses
- L12:2265 weakly suppresses
- L14:444 weakly suppresses
- L19:818 weakly excites
- L21:1056 weakly excites
- L21:1047 weakly excites
- L22:2450 weakly suppresses
- L20:2485 weakly excites
- L14:2447 weakly suppresses
- L18:1481 weakly suppresses
- L18:1564 weakly suppresses
- L4:229 weakly excites
Feeds later layers
- L27:955 strongly suppresses
- L25:1792 strongly suppresses the modal word 'may'
- L25:2051 strongly suppresses
- L24:766 suppresses
- L26:466 suppresses
- L25:1767 weakly suppresses
- L24:1814 weakly suppresses
- L27:1763 weakly suppresses
- L26:2632 weakly excites
- L25:1211 weakly suppresses
- L27:500 weakly suppresses
- L27:2787 weakly excites
- L25:2788 weakly suppresses
- L27:1130 weakly excites
- L25:1010 weakly suppresses
- L25:1294 weakly suppresses
- L26:1425 weakly suppresses
- L25:1328 weakly excites
- L26:2325 weakly suppresses
- L25:752 weakly suppresses
- L24:378 weakly excites
- L25:2063 weakly suppresses the word before 'risk'
- L27:2796 weakly excites
- L27:1571 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
- spir
- oleon
- ffd
- bie
- itals
- SPAN
- 换了
- द
- Esp
- mpr
Suppresses
- possible
- possible
- Possible
- Possible
- posible
- possível
- possibile
- 可能
- possibility
- _possible