'as' in 'as far as' and 'as many as'
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. 24 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 reacts to the word "as" in comparison phrases such as "as far as", "as early as" or "as many as". Its negative side fires on words like "over" in "overkill" and "too" in "too light-handed", which are less clearly related. When turned up or down, the model's writing became repetitive or off topic, so steering gives no clear sign.
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.
-
anches, previously the Bishops of Coutances. Standing 80 metres (295 ft) tall, it dominates the town and can be seen from as far away as the island of Jersey. It is a classic example of the Gothic
-
they don't do as well as younger people on complex memory or learning tests. Scientists have found, though, that given enough time, healthy older people can do as well as younger people do on these tests. In fact, as they age,
-
besides an exposition of Dioscorides, a Natural History dealing with all the plants known to Mattioli.<|endoftext|>The History of the hall can be traced to as far back as the Domesday Survey. With little knowledge of the hall prior
-
possible to preempt<|endoftext|>Food for thought⏎Sit down for a meal at the table, and you'll probably find foods from a half-dozen countries spread across nearly as many continents, on your plate.⏎Supermarket shelves today are likely to be stocked
-
of HCV infection is dependant upon on geography, alcohol use, viral characteristics ( different genetic types, viral load ), co-infection with other viruses and some as yet unidentified factors.⏎After exposure to the virus, detectable viral genetic material called
-
-dial or a garden bed or the residence itself, or a brook falling down the sward between trees and bushes and clumpy growths. There are as many forms and kinds of gardens as there are persons who have gardens […].”
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: how to use a perking on a perking on a perking on a perking... perking perking... perking on a perking perking... on a perking on a perking... on a perking perking
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
the Clean Water Act, it just brings a new truckload of restrictions for the people who do most to protect our water.⏎The Clean Water Restoration Act is regulatory overkill. It is written<|endoftext|>Feeling anxious about stressful events is a normal response
-
. To guarantee Southerners their much-desired states' rights, the federal government had no authority to levy protective tariffs, make internal improvements, or overrule state court decisions, while states had the right to sustain their own armies and
-
are being applied.⏎“If fracking is not done well it can have significant environmental impacts including polluting water and triggering earthquakes.⏎“I am also concerned that regulation may be too light-handed, particularly if fracking opens the door to a large-scale and widespread oil
-
y often criticized as overabundant and extravagant. Mercedes Maroto Camino (1995
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
- L13:2250 excites
- L21:2777 excites
- L22:151 weakly suppresses
- L16:628 weakly suppresses
- L18:2374 weakly suppresses
- L9:2431 weakly excites
- L10:103 weakly excites
- L19:569 weakly excites
- L2:929 weakly suppresses
- L23:206 weakly suppresses
- L7:795 weakly suppresses
- L8:1924 weakly suppresses
- L1:2164 weakly suppresses
- L11:2999 weakly suppresses
- L7:1754 weakly excites
- L4:1931 weakly excites
- L10:644 weakly suppresses
- L23:3034 weakly excites
- L18:102 weakly suppresses
- L9:2760 weakly suppresses
- L8:893 weakly excites
- L7:1938 weakly suppresses
- L4:1698 weakly excites
- L22:1723 weakly suppresses
Feeds later layers
- L26:1471 weakly excites
- L27:2822 weakly excites
- L27:1233 weakly excites
- L25:2727 weakly excites
- L26:268 weakly suppresses
- L26:890 weakly suppresses
- L26:761 weakly suppresses
- L27:991 weakly suppresses
- L25:892 weakly excites
- L26:1810 weakly excites
- L26:655 weakly suppresses
- L27:2653 weakly excites
- L25:1154 weakly excites
- L27:837 weakly excites
- L27:1458 weakly excites
- L27:1437 weakly excites
- L27:1821 weakly excites
- L25:1357 weakly suppresses
- L25:2618 weakly suppresses
- L25:2356 weakly excites
- L25:714 weakly suppresses
- L25:1863 weakly excites
- L27:888 weakly excites
- L26:2874 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
- PLY
- etus
- urses
- Cena
- ряд
- oted
- tern
- 酡
- ENCY
- voke
Suppresses
- bs
- on
- per
- quee
- len
- amb
- alsa
- Ste
- king
- Ste