comma before a trailing 'with' clause
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
The neuron fires on a comma that ends a full clause in news-style writing, where the sentence is about to continue with a phrase introduced by 'with'. When turned up, the model just repeated the word 'with' 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.
-
wave in northern Europe was made twice as likely by climate change, according to a preliminary analysis.⏎Temperatures have soared over much of Europe over the last month, regularly exceeding 30
-
than occupants of cars, and five times more likely to be injured.⏎Helmet usage is also on the decline, dropping from 66 percent of motorcyclists wearing helmets in 20
-
costs businesses approximately $7 billion per year due to the nearly 111 million work days lost. Norovirus cases are also on the rise this season, and the CDC reports that foodborne norovirus illness accounts for
-
.⏎The first step in winning the lottery is to purchase a ticket, which can be done online or at any store that sells state-approved tickets. Ticket prices vary, but most are inexpensive. Some are scratch-off cards, while others are more expensive
-
infections.⏎Sub-Saharan Africa remains most severely affected, with nearly 1 in every 25 adults living with HIV and accounting for
-
ish meat, and fish. Participants were ranked by DASH score into five groups, or quintiles.⏎The researchers found that higher DASH scores were associated with higher
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: if the main part of the world is a web of interconnected systems, such as the internet, then the main elements are all interconnected, and the main elements are not just one particular system. So if the main part is a web of systems
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
our Patrons Cristina Magistrali, Toren Wallengren, Eric Huffman, Julia Casey, Colton Siefker, Daniel Wenger, Matias Mancini, John Thompson, Chris Krish, and Alvero Wiggins for supporting
-
the second method is employed for large a.c. generators⏎!! OOPS Login [Click here] is required for more results / answer<|endoftext|>By: Swapnil Katiyar⏎With the 17th Lok Sabha elections round the corner,
-
much” is the underlying philosophy of a number of innovative programmes being carried out in Uruguayan schools that are using gardens as a teaching resource, explained Edith Moraes, director of the national Primary Education Board.⏎One of these initiatives is
-
x 66 mm|⏎|Printer||American Bank Note Company (U.S.A.) – ABNC|⏎|Signatures||Ernesto (Che) Guevara Serna & Rolando Díaz Astaraín|⏎|Comment
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
- L19:1036 suppresses
- L20:1832 suppresses 'as' introducing a reason
- L19:1714 suppresses
- L18:1530 suppresses
- L20:476 excites commas that lead into a continuation
- L22:992 weakly suppresses
- L18:1514 weakly suppresses
- L6:2407 weakly suppresses
- L20:1043 weakly excites
- L22:1489 weakly excites
- L14:1634 weakly excites
- L13:2296 weakly suppresses
- L19:612 weakly excites
- L22:2630 weakly excites
- L9:1541 weakly excites
- L2:215 weakly suppresses
- L18:753 weakly suppresses
- L7:1471 weakly suppresses
- L10:1638 weakly suppresses
- L18:366 weakly excites comma inside numbers and dates
- L11:1047 weakly excites pieces of software and library names
- L17:1537 weakly excites
- L16:846 weakly excites comma opening an aside or explanation
- L8:1366 weakly excites
Feeds later layers
- L24:921 strongly excites
- L26:741 strongly excites
- L27:603 strongly excites
- L24:883 strongly excites
- L27:1443 excites
- L26:223 suppresses
- L27:2153 excites
- L25:2139 suppresses capital N starting an abbreviation
- L24:2419 weakly suppresses
- L24:1002 weakly excites
- L25:850 weakly suppresses
- L25:2803 weakly excites
- L24:394 weakly suppresses
- L24:360 weakly excites
- L25:1564 weakly excites word pieces before 'trans' or 'transfer'
- L25:230 weakly excites
- L27:2862 weakly suppresses
- L24:1593 weakly excites
- L27:1869 weakly suppresses
- L27:1815 weakly suppresses
- L24:670 weakly suppresses
- L24:1666 weakly excites
- L26:2062 weakly suppresses
- L26:973 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
- with
- with
- With
- With
- _with
- →with
- .with
- avec
- WITH
- -with
Suppresses
- haled
- edar
- ewing
- ода
- 妫
- ül
- .createFrom
- urgeon
- iedo
- 并对