digits inside years like 19xx
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. 11 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 lights up on the digits inside years, especially the 9 in dates like 1997 or 1909. It goes strongly negative on large round numbers and ancient dates such as 15000 BC or 10,000 years ago. Turning it up made the model's text drift into unrelated rambling about the weather, the news and everyday topics.
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.
-
or Arcturus – nearer than ten parsecs1 and only about four main sequence A stars.⏎– Dutch astronomer Willem Jacob Luyten (1899 - 1994)⏎The closest star to us is,
-
and not by birth. The nobleman is a righteous individual, an example of filial piety, humane behaviour, virtue and propriety (Ren 1997, pp. 19-20; Lee Dian Rain
-
. On<|endoftext|>Anatoly Lunacharsky On Literature and Art 1909⏎The greatest of all great heroes of action was the Greek Hercules,
-
supposed, that this name originated from the Sanskrit word vihara, that means "a Buddhist monastery" [Frye, 1956, p. 106 – 119]. However
-
00 km2, within which as little as 200 km2 of primary forest may remain (T. Arndt in litt. 1993). Until recently it was judged to be fairly common (Ridg
-
not overly large. The whitish-blue patches are distributed much as in Marcus's drawing of E. canguzua.⏎• Marcus, Er. (1955) Opisthobranchia from Brazil. Boletim
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 "Hokkaido" (Hokkaido) is a culture which is a culture which is a culture which is a culture which is a culture which is a culture which is a culture which is a culture which is
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
The history of art can be traced back to cave paintings of about 15000 BC. Gotell, E., Brown, S.,
-
aceae family. Here are some Butternut Squash Health Benefits.⏎Squash is among the oldest known crops dating back 10,000 years to Mexico and America. The name squash, which means uncooked or eaten
-
acious analytical framework that focuses on all textual records throughout human history, from the earliest periods of traceable communication—perhaps as early as 60,000 BCE—to the present day. At its core, it examines the material history of
-
A Paleo diet is a diet plan based on foods similar to what might have been eaten during the Paleolithic era, which dates back from about 2.5 million to 10,000 years ago.⏎Food used to be obtained
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
- L9:382 suppresses
- L10:1746 excites
- L10:1291 excites
- L9:927 suppresses
- L10:2387 suppresses
- L3:1029 weakly suppresses
- L9:659 weakly excites
- L8:2662 weakly suppresses
- L10:1923 weakly suppresses
- L3:1035 weakly suppresses
- L8:521 weakly excites
- L6:2366 weakly suppresses
- L7:2892 weakly suppresses
- L7:139 weakly excites
- L6:2811 weakly suppresses
- L9:684 weakly suppresses
- L8:485 weakly suppresses
- L6:1987 weakly suppresses
- L10:329 weakly excites
- L10:2106 weakly suppresses
- L10:2005 weakly excites
- L7:334 weakly suppresses
- L8:49 weakly excites
- L9:1460 weakly excites
Feeds later layers
- L17:2799 suppresses
- L18:2792 suppresses
- L17:2618 suppresses
- L22:1890 suppresses
- L16:1532 suppresses
- L25:2048 suppresses
- L18:2893 suppresses
- L22:2240 suppresses
- L19:882 suppresses
- L25:1956 suppresses
- L21:2998 suppresses
- L21:1714 suppresses
- L20:2251 suppresses
- L20:1522 suppresses
- L21:2658 suppresses
- L19:2030 suppresses
- L20:1782 suppresses
- L26:615 weakly suppresses
- L27:1673 weakly suppresses
- L20:1396 weakly suppresses
- L24:3040 weakly suppresses
- L22:2292 weakly suppresses opening year or date range
- L23:958 weakly suppresses
- L23:2951 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
- IAS
- ides
- 营业
- jab
- itting
- vy
- 업
- IOS
- агент
- :],
Suppresses
- Historical
- /history
- historical
- 历史
- historia
- -history
- history
- history
- History
- 起源