third digit of years and numbers
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. 21 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 middle zero in years and long numbers, like the 0 in 2005 or 1690. Where it works in reverse, it is more about words like 'root' and 'core', and turning it down made the writing repeat 'root' and 'core' over and over.
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.
-
aimed at the asteroid 2005 YU55 a few days ago, when it passed the Earth at
-
Jacques Bovedere, said the possible applications of the Magnifier tube in the astronomical field in letter to Galileo Galilei, and in 1609 Galileo built several glasses without having seen models Dutch, using a convex lens
-
nearshore areas in French Polynesia, determined using ultrasonic telemetry⏎Centre IRD de Montpellier, BP 5045, 34032 Montpellier cedex 1, France⏎2 Centre IRD
-
very bright stars with a view for follow up observations by other telescopes.⏎Kepler, which cost $600 million, was launched in 2009. It is one of the most successful missions in NASA’s history, having
-
the categories of the three majors to be your treatments.)<|endoftext|>During this week's close encounter with Earth, astronomers tracked the aircraft-carrier-sized asteroid 2005 YU55 with radar instruments, infrared cameras, visible-light telescopes
-
the recognition of<|endoftext|>The association between British colonial rule and lower levels of corruption is often emphasized in legal origins literature (La Porta et al., 2008). However, given the historical context of Africa, we hypothesize that Britain
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: at the root of the core of the root of the root of the root of the root of the root of the root of the root of the root of the root of the root of the root of the root of the root of the root
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
.⏎Blockchain and Decentralization⏎Blockchain technology burst onto the scene with the advent of cryptocurrencies like Bitcoin. However, its applications extend far beyond digital currencies. At its core, blockchain is a decentralized and immutable ledger that can record transactions and information in
-
employees to wear it, many facilities<|endoftext|>The Israel-Palestine conflict – often referred to as the ‘world’s most intractable conflict’ – is rooted in a dispute over land claimed by Jews as their biblical birthright and by the
-
do with science. But the journey from evidence to policy is far from simple, always lengthy, frequently divisive and often ineffective. Why is that? To get to the bottom of this question we first need to ask another — what is evidence?⏎Evidence
-
religion, but the most important of the tenets. It is the very soul and the base of all the Islamic ideas and beliefs. We can say that the roots, as well as the branches of Islam take
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
- L19:2094 excites
- L9:1632 excites
- L19:2298 weakly suppresses
- L16:2820 weakly excites
- L14:985 weakly excites
- L8:2915 weakly excites
- L6:90 weakly suppresses
- L5:1794 weakly excites
- L16:2082 weakly suppresses
- L8:2120 weakly excites
- L19:1545 weakly excites
- L18:222 weakly excites
- L19:2520 weakly excites
- L19:3012 weakly excites
- L17:2653 weakly suppresses
- L4:702 weakly suppresses
- L17:2716 weakly excites
- L13:2890 weakly excites "still" or "as long as" meaning continuing
- L18:1497 weakly excites
- L15:2807 weakly suppresses
- L7:30 weakly suppresses
- L14:1057 weakly excites
- L19:2486 weakly suppresses
- L11:246 weakly suppresses
Feeds later layers
- L26:1543 suppresses
- L23:2966 suppresses
- L25:2036 suppresses
- L27:2290 suppresses
- L22:1612 suppresses
- L25:2502 suppresses
- L24:2635 suppresses
- L25:2040 suppresses
- L27:1376 suppresses
- L22:2708 suppresses
- L25:1980 suppresses
- L22:1701 weakly suppresses
- L23:2470 weakly excites
- L22:3052 weakly suppresses
- L22:927 weakly suppresses
- L24:2468 weakly suppresses
- L27:2037 weakly suppresses
- L22:1344 weakly excites
- L24:2970 weakly excites
- L23:2529 weakly suppresses
- L27:1327 weakly suppresses
- L26:2488 weakly suppresses
- L25:942 weakly suppresses
- L22:1172 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
- migrationBuilder
- secutive
- ivities
- 过得
- 平均
- ặp
- 掖
- ulia
- اة
- icopter
Suppresses
- root
- 根
- 根本
- 根基
- root
- Root
- Root
- fundamental
- roots
- core