the spot right before a year's last digits
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
The neuron fires on the blank spot just before the last digits of a year, such as the gap in 'from 1990 to ___2010'. It is strongly negative on the first digit '2' of years like 2001 or 2020. Turning it up gave unrelated text, and turning it down gave repeated 'zero' words.
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.
-
in school to increase from 75 to 88 percent. MENA achieved the same increase in 2 decades: from 1990 to 2010.2⏎However, it is important to note that regional
-
1994) and at Cadamosto⏎seamount (Heleno et al 1999; Grevemeyer et al. 2010). The rectangle shows the position of Fig. 1b
-
1912-2006). Among others that he strongly influenced was the U. S. Federal Reserve Chairman from 1987 to 2006, Alan Greenspan.⏎In 1999 Fre
-
isolation of small populations are major concerns in the conservation of large carnivores (Clark et al., 1996; Singleton & Lehmkuhl, 2001). The loss of habitat and connectivity between populations, compounded by
-
D member countries. The STEMG was originally formed in 1983 as a working group to develop AECMA Simplified English and subsequently, in 2004, when ASD was created,
-
every few days (Salaman and Giles 1995, P. Boesman in litt. 1998, Snyder et al. 2000) on the relatively well-watched San Lorenzo Ridge. It
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:000000000000000000000000000000000000000000000000
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
change. Even by California standards; it’s ambitious, calling for a 36% reduction in greenhouse gas emissions from 2005 levels by 2020 (statewide emissions decreased 5.3% between 2
-
e-learning market in the United States is expected to grow $6.8 billion by 2015, up from $2.9 billion in 2010. According to Sanjeev Ahuja, the vice president of
-
all that remains of the giant Buddhas carved into the cliffs are the indentations that once housed them. The Taliban blew these figures up with dynamite in 2001, viewing them as idols from a past they despise
-
the issue by modifying the ALA<|endoftext|>Note: I published a previous version of this article on The Change Blog (now called Possibility of Change) in 2011.⏎What is life? Is it the state of physical existence,
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
- L21:2497 strongly suppresses
- L20:1520 excites
- L19:2444 excites
- L20:2472 excites
- L20:1607 excites
- L21:2903 suppresses
- L21:741 suppresses
- L17:2118 suppresses
- L11:2530 excites
- L17:1476 excites
- L0:305 weakly excites
- L20:1355 weakly suppresses
- L18:2103 weakly suppresses
- L17:1224 weakly excites
- L12:697 weakly suppresses
- L12:2691 weakly suppresses
- L22:906 weakly suppresses
- L17:1416 weakly suppresses
- L18:1692 weakly suppresses
- L13:2285 weakly excites
- L20:919 weakly suppresses
- L6:1747 weakly excites technical word fragments, splicing and isotopes
- L22:504 weakly excites
- L19:234 weakly suppresses
Feeds later layers
- L25:473 strongly suppresses zeros inside large round numbers
- L24:2703 strongly suppresses
- L27:153 strongly excites the spot just before a historical number
- L26:413 strongly suppresses
- L27:862 suppresses
- L27:189 suppresses
- L27:1153 suppresses
- L25:92 suppresses
- L25:2960 excites
- L25:543 excites
- L27:396 suppresses
- L27:992 excites
- L27:352 excites
- L26:834 excites
- L27:294 weakly suppresses
- L26:682 weakly excites
- L27:1202 weakly suppresses
- L25:2314 weakly excites the 1 starting a 1900s year
- L25:1895 weakly suppresses first digit of a year like 2020
- L25:379 weakly excites
- L24:969 weakly excites
- L24:1318 weakly suppresses
- L26:137 weakly excites
- L25:2047 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
- bidden
- pecially
- 上世纪
- OURS
- ieties
- ilitation
- urally
- ictions
- емся
- ({...
Suppresses
- 0
- ₀
- orio
- 0
- zero
- �
- ��
- kie
- aine
- -zero