the 1 starting a 1900s year
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. 25 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 "1" that starts a year from the 1900s, such as the "1" in 1979 or 1993. Turning it up made the model's writing fill with nines, which is only loosely related.
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.
-
MBCT's main technique is based on the Mindfulness-Based Stress Reduction (MBSR) eight week program, developed by Jon Kabat-Zinn in 1979 at the
-
Day? Be like Phil the weatherman, and try some mindfulness⏎Many of us will recall the American comedy film "Groundhog Day."⏎Originally released in 1993, it stars the incomparable Bill Murray as Phil Conners,
-
chip to replace entire rooms full of computers (Fig. 1).⏎The first microprocessor was the Intel 4004 created in 1971 and was made primarily for use in calculators . By today's
-
step was to build a device to read the traffic tapes directly and eliminate the tedious manual work. The Intel 8008 microprocessor was announced in 19
-
a popular method for training neural networks. For example, one of the early papers in that area “Training Feedforward Neural Networks Using Genetic Algorithms” published in 1989, has been cited more than 1200 times.
-
youth to the violent death of Lennon's mother during his adolescence, from Lennon's post-Ed Sullivan skepticism to John and Yoko's Bed-Ins of 1969 - and unearths the secrets of its lasting import. If
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 backos, backos, backos, backos, backos, backos, backos, backos, backos, backos, backos, backos, backos, backos, backos, backos
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
about a “bomber gap” and later a “missile gap.” On October 4, 1957, Sputnik, the Soviet satellite that was the first to orbit Earth, shocked Americans and delighted the Soviets.
-
launch "gave the program a bad reputation that I don't think it really deserved," says Neufeld. "The Vanguard launch came after the success of Sputnik 1 and Sputnik 2,
-
who maintained that Germany did not join the West until after Auschwitz (1).⏎This diagnosis was reworked by the historian François Furet after the fall of the Soviet Union and the end of the cold war. Furet interpreted Nazism and
-
10.aug.25 » We have the Russians to thank for the Internet. In 1957 they won the race into space by launching Sputnik (satellite with a cool name), which remained in orbit for some
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
- L21:1815 excites
- L20:642 suppresses
- L23:1369 suppresses the '0' in years like 2008
- L18:2542 excites
- L2:111 suppresses
- L21:453 weakly excites
- L23:1243 weakly excites the spot right before a year's last digits
- L20:1954 weakly suppresses
- L20:1315 weakly suppresses
- L20:529 weakly excites
- L23:3058 weakly excites
- L16:532 weakly excites
- L9:1604 weakly suppresses
- L12:1251 weakly suppresses
- L20:1396 weakly excites
- L8:323 weakly suppresses
- L6:1183 weakly excites
- L2:75 weakly suppresses
- L19:1583 weakly excites
- L8:2018 weakly suppresses
- L0:1179 weakly excites
- L24:875 weakly suppresses
- L20:124 weakly excites
- L8:710 weakly suppresses
Feeds later layers
- L27:893 strongly excites
- L27:494 suppresses
- L27:1087 excites
- L26:2492 excites
- L27:628 weakly excites
- L26:1248 weakly suppresses
- L26:580 weakly suppresses
- L27:1519 weakly suppresses
- L26:528 weakly excites
- L26:1437 weakly excites
- L27:2600 weakly suppresses
- L26:1334 weakly excites
- L26:2165 weakly suppresses
- L27:1232 weakly suppresses
- L27:1268 weakly suppresses
- L26:1339 weakly excites
- L26:905 weakly excites
- L27:2120 weakly suppresses
- L26:1966 weakly excites
- L27:1209 weakly excites
- L27:1502 weakly suppresses
- L27:1606 weakly excites
- L27:2530 weakly excites
- L26:1158 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
- 9
- 九
- nine
- 9
- ninety
- Nine
- 九十
- -nine
- Nine
- nine
Suppresses
- 族自治县
- ocities
- stile
- 膦
- .present
- 蘇
- atives
- ежду
- ommen
- 共有