second digit of years like 2020
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 "0" in the year 2020 or 2021 when it comes right after the "2". It also fires on years like 2002 or 2001 with a negative sign. Turning it up made the output fall into repeated "2222" strings, and turning it down gave repeated zeros.
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.
-
are. You may even have heard of names like Bitcoin or maybe Ethereum to name a view. There are many most popular cryptocurrencies which can be used in 2020.⏎Maybe you’ve heard the terms digital gold or halving if you
-
post and review on that article. https://t.co/D9NIorgXuc— Bruce Rogers (@bjrmd) August 30, 2021⏎PACING: Dist
-
their data protection measures to safeguard their data and hardware assets. As a guide, here are the five biggest IT security threats that you should prepare for as 2016 comes.⏎1. Internet of Things (IoT) security⏎The
-
healthcare industry today is the most targeted industry in terms of cyber security threats. A recent study by Critical insights (A cyber security company) shows that in 2021 alone a whopping number of 45 million individuals were affected by cyber
-
seconds or use hand sanitiser with a minimum of 60% alcohol base.⏎9. Wash<|endoftext|>Date Updated: 12/17/2020⏎Strep throat is a bacterial infection that can make your throat feel
-
households in Nigeria do not have sufficient access to soap and water for handwashing and that there is low adherence to hygiene protocols (National Bureau of Statistics, 2020; Premium Times, 2020).⏎The most recent Afro
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: The number of just-type meaning-type (0000000000000000000000000000000000000000
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
the success of the struggle for equal rights for women and minorities<|endoftext|>The No Child Left Behind Act, signed into law by President Bush January 8, 2002 includes as one of its measures the standard to staff our public schools with
-
9⏎rothbard on the utility of quantitative easing⏎Economist Murray Rothbard ("America‟s Great Depression", Ludwig von Mises Institute 2000) made extensive analysis of why the inflationary policies of the US Federal
-
2010⏎In 2009, Gulf War-era II veterans—those who had served in the Armed Forces sometime since September 2001 and had returned to civilian life—were more likely to have some college
-
local law enforcement in northern Texas and developed this early warning system to help find abducted children. National coordination of state and local AMBER Alert programs began in 2003.⏎The AMBER Alert Plan is activated when a child 17
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
- L17:1476 excites
- L14:1700 excites
- L18:1976 weakly suppresses
- L19:843 weakly suppresses
- L14:2966 weakly excites
- L21:2497 weakly suppresses
- L7:2258 weakly excites
- L11:1975 weakly suppresses
- L20:1520 weakly excites
- L2:2517 weakly suppresses
- L13:859 weakly suppresses
- L19:1145 weakly excites
- L1:416 weakly suppresses
- L16:1677 weakly excites
- L0:1071 weakly excites
- L4:2138 weakly excites
- L21:2106 weakly excites
- L12:2450 weakly excites
- L16:2238 weakly suppresses
- L8:2323 weakly excites
- L4:26 weakly excites
- L2:59 weakly excites
- L3:405 weakly excites
- L0:121 weakly excites
Feeds later layers
- L27:189 strongly suppresses
- L26:834 strongly excites
- L27:352 excites
- L25:379 excites
- L27:992 excites
- L25:1478 excites
- L27:294 suppresses
- L25:224 suppresses
- L25:473 suppresses zeros inside large round numbers
- L27:862 suppresses
- L27:153 excites the spot just before a historical number
- L25:2298 excites
- L27:396 suppresses
- L26:413 suppresses
- L24:1446 excites
- L27:1157 excites
- L26:848 suppresses
- L24:677 excites
- L27:1202 suppresses
- L27:1153 weakly suppresses
- L26:598 weakly excites spot just before a year or century
- L24:342 weakly excites unit conversions in parentheses
- L26:905 weakly suppresses
- L27:205 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
- 2
- 二十
- 二十五
- twenty
- twenty
- 二十四
- 2
- 十八
- 十七
- kır
Suppresses
- tery
- ấy
- conomy
- 0
- auté
- uisine
- луш
- tür
- ldb
- чная