'time' in 'for the first 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. 5 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 word 'time' in the phrase 'for the first time', and it is quite reliable on that phrase. Turning it up pushed the output toward 'first time' and first-ever events, so the steering matches the examples.
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.
-
970's when President Nixon visited mainland China for the first time. Today, that may not seem like such a big deal. But back t
-
96, the first modern Olympic Games are held in Athens, Greece, with athletes from 14 countries participating.⏎The International Olympic Committee met for the first time in Paris in June 1984 and chose Greece as the site of
-
It turns out that robot design could become much more than just a research tool.⏎Team leader Robert J Full of UC Berkeley states, "We showed for the first time that lizards swing their tail up or down to counteract the rotation of their
-
accelerators.⏎In 1956, a team of physicists led by Americans Clyde Cowan and Frederick Reines observed the ghostly particles for the first time. For their discovery, Reines shared the 1995 Nobel Prize
-
repeated it a few times since then.⏎But there were two new dimensions of the recent flight. One, it carried a Chinese woman astronaut into space for the first time. Secondly and far more significantly, it demonstrated China's capability to conduct docking of
-
one of largest cities of the region, also belongs to this number. This name was mentioned for the first time on the earliest copper coins of Buxārā (4th –
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 the month is a 1000-year-old tree in the middle of the 2014 cube, the number of the year is a 300-year-old tree, the number of the year is
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
criteria.<|endoftext|>The haunting image of the fallen soldier, aged just 18, is one of a hundred portraits of those who served in the First World War published on Friday by the Imperial War Museum (IWM).⏎The museum is
-
previously done laboriously by hand in hundreds of weavers’ cottages were brought together in a single cotton mill, and the factory was born. The second industrial revolution came in the early 20th century, when Henry Ford mastered the moving
-
students learn to tell and write time from analog and digital clocks to the nearest five minutes, using a.m. and p.m.⏎Common Core: 2.MD.7⏎
-
961, the 23rd Amendment let District of Columbia residents vote in Federal Elections.⏎In 1964, the 24th Amendment let vote those who could not pay a poll tax.⏎In 196
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
- L4:867 suppresses
- L4:1034 weakly excites
- L1:1993 weakly suppresses
- L0:2949 weakly suppresses
- L4:1662 weakly suppresses
- L0:321 weakly excites
- L4:2999 weakly excites
- L0:846 weakly suppresses
- L4:517 weakly suppresses
- L0:2927 weakly excites
- L0:998 weakly excites
- L2:2437 weakly suppresses
- L4:83 weakly excites
- L4:1267 weakly excites
- L2:650 weakly suppresses
- L3:2724 weakly suppresses
- L2:2303 weakly suppresses
- L4:2201 weakly suppresses
- L3:2698 weakly excites
- L0:2937 weakly excites
- L0:2213 weakly suppresses
- L2:82 weakly excites
- L1:2482 weakly suppresses
- L1:2639 weakly suppresses
Feeds later layers
- L6:2213 excites
- L18:1378 excites
- L11:1663 excites
- L6:951 weakly excites
- L21:1970 weakly excites
- L11:2781 weakly excites
- L14:1778 weakly excites
- L10:1374 weakly excites
- L19:2552 weakly excites
- L8:2064 weakly excites
- L11:1266 weakly suppresses
- L23:2112 weakly suppresses
- L18:2837 weakly suppresses
- L13:2892 weakly excites
- L8:2322 weakly excites
- L14:3026 weakly suppresses
- L8:1708 weakly excites
- L17:3018 weakly suppresses
- L16:885 weakly excites
- L6:2819 weakly suppresses
- L12:1537 weakly excites
- L27:793 weakly excites
- L18:1125 weakly excites
- L13:63 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
- ERICA
- 机会
- 等到
- les
- blush
- acles
- 美的
- orque
- sourceMappingURL
- egl
Suppresses
- jąc
- alyze
- kip
- anoia
- annya
- úde
- télé
- 隍
- 릭
- umen