year digits inside 20xx dates
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. 15 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
Fires on the digit in the middle of a year like 2020, 2021 or 2023, the one that tells the decade. Turning it up gave text that did not obviously relate to years, so the link to the pattern is weak in the output.
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.
-
Antonio vied with each other for trade, commerce and business.⏎George R. Woolfolk⏎Prairie View A&M College⏎Rights© 2021 Prairie View A & M University⏎This work is licensed under a Creative Commons
-
. So, you could also argue the Internet (<|endoftext|>このトピックには0件の返信が含まれ、1人の参加者がいます。2020年7月6日 at 9:45 AMに Megan Smith
-
This post contains my opinions and is not meant to be taken as medical advice. If you have a medical concern, please consult your personal physician.⏎@2020,<|endoftext|>The global energy system works on timescales of decades rather years.
-
the file name for working⏎Example 1: Unzipping files⏎In this example, the .tgz file of Softmaker-free office-2021 is used. let’s first check the .tgz file location and get
-
with our community members and organizations to share information with the public on the health benefits and risks linked with eating fish. See our most recent projects below.⏎2020: Asian women of child-bearing age in
-
It also offers tips for taking a social media break.⏎Stepping away from social media, even for an indefinite time, has many benefits.⏎A 2020 study found that a fear of missing out, known as FOMO,
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 have been, as I have, in the time, in the time, as it was, in the time, as it was, in the time, in the time, in the time, in the time, in the time,
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
of the European Geriatric Medicine study are available here: http://www.europeangeriaticmedicine.com/article/S1878-7649(13)00123-X/).⏎It is possible nursing staff
-
iversity and Conservation (2004) 13: 2511. doi:10.1023/B:BIOC.0000048451.07820.
-
Jake Matijevic."⏎- Curiosity on Mars sits on rocks similar to those found in marshes in MexicoTue, 30 Oct 2012, 11:07:11 EDT⏎
-
ern at gmail.com⏎Fri Sep 10 00:10:22 CEST 2010⏎On 9/9/10 4:39 PM, Baba wrote:⏎> In below code
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
- L12:1264 excites gap before a year in dates
- L10:1304 suppresses
- L3:2963 weakly excites "discrete" and "dispersive" in technical text
- L6:1130 weakly suppresses
- L2:1096 weakly excites
- L14:1700 weakly excites
- L1:1235 weakly excites
- L4:2472 weakly excites
- L11:1848 weakly excites
- L2:996 weakly excites
- L13:3051 weakly suppresses
- L9:2227 weakly suppresses ordinal numbers in military unit names
- L9:540 weakly suppresses
- L5:600 weakly excites
- L0:393 weakly suppresses
- L6:1654 weakly suppresses
- L2:2805 weakly suppresses
- L5:32 weakly suppresses
- L12:696 weakly suppresses
- L7:1554 weakly excites
- L4:2720 weakly excites
- L1:461 weakly excites
- L7:2028 weakly suppresses
- L7:399 weakly suppresses laboratory and research facility terms
Feeds later layers
- L24:1213 weakly suppresses
- L18:1368 weakly suppresses
- L17:2356 weakly suppresses
- L22:1435 weakly excites
- L18:819 weakly suppresses
- L21:2416 weakly excites
- L23:449 weakly excites
- L16:1161 weakly suppresses
- L22:2036 weakly excites
- L22:2346 weakly excites
- L18:1816 weakly excites
- L24:695 weakly suppresses
- L26:2186 weakly excites
- L20:2196 weakly excites
- L19:1753 weakly suppresses
- L22:2594 weakly excites
- L21:811 weakly excites
- L19:2130 weakly excites
- L22:1541 weakly suppresses
- L21:123 weakly excites
- L21:1678 weakly suppresses
- L24:1901 weakly suppresses
- L17:1757 weakly excites
- L17:2474 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
- adow
- werk
- Alpine
- Elk
- 峦
- recomm
- 文学
- alış
- emphasize
- fearless
Suppresses
- .au
- 期内
- held
- xAA
- vase
- ipples
- erde
- nerRadius
- >()->
- 一艘