name pieces inside a multi-part word
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. 1 of Qwen3-0.6B. Its strongest snippets are mixed: they peak on different tokens. 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 middle or second half of a name or phrase that is split up, like 'Web' in 'Amazon Web Services' or 'ville' in 'Bartlesville'. It is weak overall, and turning it up did not produce a clear theme in the output.
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.
-
sites or client-side applications. Most of these services are not exposed directly to end users, but instead offer functionality that other developers can use in their applications. Amazon Web Services’ offerings are accessed over HTTP, using REST and SOAP protocols. All services
-
. The principal producing horizon is the Bartlesville sand, found at an average depth of 1,150 feet.
-
, services, and supports. Lifetime costs to care for an individual with ASD have been estimated to be $3.2 million. ⏎With funding from the American Recovery and Reinvestment Act (Recovery Act), the U.S. Department of Health
-
from the boundaries of asia to places as far away as Spain, New Zealand and the Americas. From the Migyur Dorje Stupa in Poolesville, Maryland to the Mahabodhi Stupa in Stokes Valley, New Zealand
-
Amazon Web Services⏎|Type of site||Web service, cloud computing|⏎Amazon Web Services (AWS) is a collection of remote computing services (also called web services
-
constant air exchange regardless of weather conditions,⏎- filtration of supplied air by means of special filters,⏎- savings on heating as a result of recovering heat<|endoftext|>In the beginning, they rallied behind Hitler in the national interest of Germany; in the end,
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: with the to. it. and. and. and. and. and. and. and. in. and. and. in. and. and. and. and.. and. for. and. up. and..
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
of giant ice made up of frozen nitrogen mixed with carbon monoxide and methane.⏎It covers an area four kilometers deep and a thousand kilometers wide in the Sputnik Plain of Pluto.⏎To understand the origins of this
-
reputation that I don't think it really deserved," says Neufeld. "The Vanguard launch came after the success of Sputnik 1 and Sputnik 2,
-
.aug.25 » We have the Russians to thank for the Internet. In 1957 they won the race into space by launching Sputnik (satellite with a cool name), which remained in orbit for some 3
-
, has been unveiled in the Oeiras municipality in the Lisbon metropolitan area, the Russian embassy in Portugal told Sputnik.⏎“We will always remember the names of the first daredevils who made the
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
- L0:1301 suppresses
- L0:1224 suppresses
- L0:178 suppresses
- L0:283 excites
- L0:1817 suppresses
- L0:851 weakly suppresses
- L0:1279 weakly excites
- L0:305 weakly excites
- L0:794 weakly suppresses
- L0:2093 weakly excites
- L0:632 weakly suppresses
- L0:1845 weakly suppresses
- L0:1940 weakly suppresses
- L0:1730 weakly suppresses
- L0:84 weakly suppresses
- L0:2599 weakly suppresses
- L0:1043 weakly excites
- L0:2634 weakly suppresses
- L0:591 weakly suppresses
- L0:2303 weakly excites
- L0:2124 weakly excites
- L0:2009 weakly suppresses
- L0:1069 weakly suppresses
- L0:1329 weakly suppresses
Feeds later layers
- L2:85 suppresses
- L3:94 suppresses
- L2:642 suppresses
- L2:325 suppresses
- L4:367 suppresses
- L2:302 suppresses
- L2:670 suppresses
- L2:1566 suppresses
- L2:93 excites
- L2:913 excites
- L3:676 suppresses
- L6:819 excites
- L2:427 excites
- L3:529 excites
- L4:218 excites
- L3:311 suppresses
- L2:1504 suppresses
- L5:305 suppresses
- L3:320 weakly suppresses
- L5:2680 weakly excites
- L2:1750 weakly suppresses
- L6:87 weakly suppresses
- L2:1467 weakly suppresses
- L19:1667 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
- sätze
- periment
- эм
- 'à
- filer
- 娶
- ramid
- ��
- ласт
- pler
Suppresses
- .HandlerFunc
- serviceName
- angelo
- Egypt
- eking
- RequiredMixin
- ittel
- izio
- ence
- acco