time of day before a clock time
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
This neuron fires on the empty spot right after words like "at", "to", or "from" where a clock time is about to come, such as "open until 4:30 p.m.". Turning it up gave text that repeated words, so the steering results are not a clear match.
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.
-
.⏎Visit our Ed Clark Museum of Missouri Geology on Geologic Map Day to learn more about geologic maps. Open weekdays from 8 a.m. to 4:30 p.m., we are located at 111 Fair
-
coding is waiting for you, and it fits in the palm of your hand.⏎Have an amazing project to share? The Electronics Show and Tell is every Wednesday at 7pm ET! To join, head over to YouTube<|endoftext|>Definition of Parab
-
to allow its black patrons to eat at its lunch counter. Though they received no service, the four men sat quietly and without incident. When the store closed at 5:30, they left.⏎The next morning, the four young men
-
theoretical possibility of space flight, but built and flew the world’s first liquid propellant rockets.”⏎The event will take place from 10 a.m. to 3 p.m. on Saturday, June 11, on the Mall in
-
for each session. The camp will meet daily Monday – Friday from 9am – 4pm.<|endoftext|>People with diabetes are at increased risk of
-
am sure you are all raring to go!!⏎I am going to set up a 'Zoom' meeting on Thursday 14th May at 12pm as I have had lots of emails from you and your parents saying
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: what if we have a random family of random variables that are all independently and independently randomly randomly randomly random. I need to solve this for a random family of random variables. I think I have a correct family of random variables that satisfies this,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
and I wrote this document.⏎QRcode is developed by DENSO Corporation. It is standardized by JIS and ISO.⏎This document is based on JIS X 0510(1999).⏎If you want to know more
-
only mathematicians - like to consider more specific questions in order to get some feeling for what is happening. In this article we will reflect about a special class of Lunda-designs.⏎On the day before my daughter Likilisa's fourth birthday
-
engagements. During World War II, four battles took place for control of the city: The city was captured and recaptured by Nazi Germany on 24 October 1941; there was a disastrous Red Army offensive that failed to capture
-
provider:|| Postal: Katajanokanlaituri 6B, 00160 Helsinki|⏎Web page: http://www.wider.un
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
- L21:128 excites
- L3:461 excites
- L20:293 suppresses
- L21:1328 weakly suppresses
- L21:2349 weakly excites
- L16:566 weakly suppresses unclear: mid-sentence word choices, mixed
- L9:693 weakly suppresses
- L16:2997 weakly excites
- L4:739 weakly suppresses
- L0:1277 weakly excites
- L10:276 weakly excites
- L10:2435 weakly excites
- L4:1681 weakly excites
- L22:1048 weakly suppresses
- L13:1479 weakly suppresses
- L10:2217 weakly suppresses
- L10:455 weakly excites
- L9:65 weakly suppresses
- L3:531 weakly excites
- L4:1542 weakly suppresses
- L10:1458 weakly excites
- L15:1616 weakly suppresses word fragments mid-word
- L19:1103 weakly suppresses
- L18:408 weakly excites
Feeds later layers
- L27:2522 weakly excites
- L24:2837 weakly excites
- L24:1562 weakly suppresses
- L25:324 weakly excites
- L26:1234 weakly suppresses
- L25:1709 weakly excites
- L27:1320 weakly excites
- L25:2735 weakly excites
- L26:2522 weakly suppresses
- L24:1066 weakly excites
- L25:2792 weakly excites
- L24:627 weakly excites
- L25:2444 weakly excites
- L25:1346 weakly excites
- L25:124 weakly excites
- L24:2531 weakly excites
- L24:293 weakly excites
- L24:1041 weakly suppresses
- L24:2703 weakly suppresses
- L26:1263 weakly excites
- L24:2310 weakly suppresses
- L27:1311 weakly suppresses
- L25:1953 weakly suppresses
- L24:2445 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
- measure
- جه
- 纥
- mes
- .Measure
- bek
- andin
- regulation
- isper
- inar
Suppresses
- 穿梭
- nage
- tastes
- apsed
- randomly
- herits
- entious
- [random
- giấc
- apers