the 'per' piece in perished and per se
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. 25 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 lights up on the small word piece 'per' when it starts words like 'perished', 'per se' and 'permissive'. Turning it up made the model write confused, repetitive text, so the steering results don't reveal a clear theme.
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.
-
a time reinforced by soldiers, were kept at bay for many hours by two foreign burglars who defended themselves in a house with Mauser pistols, and who ultimately perished when the building caught fire and was burnt.⏎The following table shows the war
-
cohesiveness while living in a Christian world. It will be suggested that Mudejar self-perception and group-perception were founded not only on religious belief per se, but also on⏎the perpetuation of a social world the organization and
-
to the elimination of the numbered posts and the majority win feature, because as a practical matter women were no longer seen as competing against men or incumbents per se. Moreover, I see the system as giving women an excellent opportunity to organize
-
basic board with future potential built in - you can buy Arduino shields to do a wide variety of things - it doesn't have an Operating System (OS) per se. Arduino has a great deal of support for accessing the board and driving serv
-
28 "And you shall say to them, 'This is the nation that did not obey the voice of the LORD their God or accept correction; truth has perished and has been cut
-
ic Buddhist Apologetics (PDF)⏎Antinomianism comes in at least two forms: strong or weak. Weak antinomianism is permissive: one may do what is bad and nevertheless somehow remain not guilty. Strong
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 make a single single string, but not through a single string, but through a single string, but through a single string, but through a single string, but through a single string, but through a single string, but through a
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
iking Trails for Dogs||5 Ways to Get Your Cat with Mobility Issues Moving||5 Ways to Get Your Dog with Mobility Issues Moving||5 Exercise Tips for Arthritic Cats|⏎|Top Ten 'Small Breed' Dogs||Top Ten Dogs for
-
990 to 2019. The researchers analyzed life expectancy data for the 3 largest race/ethnicity groups by state: Hispanic, non-Hispanic Black, and non-Hispanic White Americans. They found that although
-
face of a threat. Birds raised in an urban environment react differently than country-bred birds when faced with a predator.⏎A study undertaken by Diego Ibanez-Alamo, researcher at the University of Granada (UGR) and Anders P
-
Dionysius II 65: οὔτε γὰπ τὸ χωρίον τοῦτο, ἐν ᾧ τὸ ἰερὸν φυ
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
- L22:530 suppresses
- L21:1934 suppresses
- L17:257 suppresses
- L20:1120 excites
- L24:105 weakly suppresses
- L19:1351 weakly suppresses
- L13:363 weakly suppresses
- L7:2329 weakly excites
- L22:366 weakly excites
- L7:1982 weakly suppresses
- L16:2974 weakly excites
- L8:2295 weakly excites
- L18:811 weakly excites
- L21:237 weakly suppresses
- L14:178 weakly suppresses
- L21:2715 weakly suppresses
- L21:2323 weakly excites
- L2:265 weakly suppresses
- L16:1754 weakly excites
- L13:2346 weakly suppresses
- L14:373 weakly suppresses
- L18:986 weakly excites
- L19:2348 weakly excites
- L15:12 weakly excites
Feeds later layers
- L27:2799 weakly suppresses the letter d inside words
- L26:3066 weakly excites
- L26:294 weakly excites
- L26:2936 weakly excites
- L27:915 weakly suppresses unclear: mixed word fragments and phrases
- L27:3014 weakly suppresses
- L26:2980 weakly suppresses
- L27:2308 weakly suppresses
- L27:1074 weakly suppresses
- L27:919 weakly suppresses
- L27:1515 weakly excites
- L26:1523 weakly excites
- L27:526 weakly excites
- L27:1426 weakly suppresses
- L27:41 weakly suppresses
- L27:501 weakly suppresses
- L26:2392 weakly suppresses
- L27:1105 weakly suppresses
- L26:2338 weakly suppresses
- L26:466 weakly excites
- L27:263 weakly suppresses
- L26:930 weakly excites
- L26:1688 weakly excites
- L27:146 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
- 真的
- лаг
- ля
- 真
- ไหล
- carg
- ham
- дав
- grátis
- htar
Suppresses
- alytics
- isations
- ��
- .appspot
- 消
- /array
- resenter
- eding
- /course
- conde