'of' after 'issue' in citations
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. 7 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 is weak and fires on the word "of" in citations like "the January issue of Neuropsychology." It only reaches modest strength, so it is a minor detail. Turning it up made the model write unrelated text about studies and plans, so the effect is hard to read.
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.
-
the word for pheasant.<|endoftext|>North American freshwater fishes are going extinct at an alarming rate compared with other species, according to an article in the September issue of BioScience. The rate of extinctions increased noticeably after 1950
-
181 calories in the same amount of time running at a pace of 5 miles per hour. A study in the May 1996 issue of the "Journal of the American Medical Association" also concluded that running on a treadmill
-
important behaviors such as remembering to take medicine at a certain time or getting to a doctor’s appointment. The research appears in the January issue of Neuropsychology, which is published by the American Psychological Association.⏎People with this
-
according to an article in the January 2006 issue of AGD Impact, the newsmagazine of the Academy of General Dentistry
-
⏎By: Ken Klein, Cosmetech Labs⏎Posted: February 1, 2010, from the February 2010 issue of Cosmetics & Toiletries.⏎Editor’s note: Cosmetics & Toile
-
light to 38 miles per hour by shooting a laser through an extremely cold gas of sodium atoms, as reported in the February 1999 issue of Nature.⏎“We have really created an optical medium with crazy, bizarre properties,” said
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: I am not a member of the group, and I have been thinking about how to help others. I am not in the group, and I have been thinking about how to help others. I am not in the group, and I have
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
chance of failing to pass on a beneficial mutation ).⏎As noted elsewhere (see ‘Giving up on reality’), Lenski seemed to have given up on ‘evolution in the
-
of the difference between 'wearing' and 'scratching.' Wearing is where the glazed surface
-
many coins bear the likeness of Herod the Great?⏎Given that Herod the Great (37–4
-
your children and in turn, they will be kind to others. It is a amazing what a good message these gestures will send to them.⏎Teach The Value of Money⏎Yes, kids see you passing these green pieces of paper around, but
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
- L6:296 suppresses
- L6:1752 excites
- L6:1741 excites
- L6:108 excites
- L6:228 excites comma after a filler word
- L6:429 suppresses
- L6:1534 suppresses
- L6:1031 suppresses
- L6:858 excites
- L4:16 excites
- L6:589 excites
- L6:436 excites
- L6:1582 suppresses
- L5:418 suppresses
- L6:325 excites
- L6:388 excites
- L6:1105 excites 'that' after a finding or surprise verb
- L6:1424 excites
- L6:592 excites
- L4:1891 excites
- L6:158 excites
- L6:98 suppresses
- L6:439 excites
- L5:2828 excites
Feeds later layers
- L8:59 suppresses
- L10:40 suppresses
- L9:330 excites gap where a number was removed
- L9:459 excites
- L9:141 suppresses
- L9:14 suppresses
- L8:615 excites
- L17:352 suppresses
- L8:390 excites "a" after sentence-opening "As"
- L11:55 weakly suppresses
- L9:197 weakly suppresses the blank spot right before a number
- L18:63 weakly excites
- L9:222 weakly excites
- L14:27 weakly suppresses 'says' after a quote
- L25:16 weakly suppresses
- L18:670 weakly excites
- L19:93 weakly suppresses
- L10:459 weakly excites
- L8:200 weakly excites
- L17:935 weakly suppresses
- L9:1412 weakly suppresses the word "be" after modals and "to"
- L9:316 weakly excites
- L9:52 weakly suppresses
- L16:745 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
- ArgumentNullException
- собой
- 濠
- 了我的
- 的风险
- 들의
- licit
- ように
- esco
- 一道
Suppresses
- fibr
- ,nonatomic
- inea
- devour
- tones
- Table
- ibia
- frequ
- 吸取
- ώ