comma inside large 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. 3 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 a comma sitting inside a big number, like the comma in "$157,000" or "6,000 lbs". Turning it up made the writing repetitive and looping, which does not show much about the comma itself.
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.
-
Lubantuun in 1927.”⏎The Crystal Skull is often referred to as the “Skull of Doom”; the skull was made from a single, unusually large block of clear
-
Curtiss and purchase a plane for $7,000 ($157,000 in today’s money). While waiting for it to be built,
-
4 wheels required to carry a 6,000-lb load would use the following calculation:⏎[(4,000 lbs + 6,000 lbs) × 1.25] ÷ 4
-
14 Star Tribune<|endoftext|>For practical purposes, the M1 Ultra acts like a single, impossibly large slice of silicon that does it all. Apple’s most powerful chip
-
of effectiveness) each year produces over $350,000 more income for his/her students compared to an average teacher. On
-
the commands you give and the M-file are somewhat disparate. The process is illustrated using a single, simple differential equation:⏎Integrate this equation from t = 0 to t =
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 "b" and "c" in the "b" and "c" market. It is in the "b" and "c" market. It is in the "b" and "c" market. It is in
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
into what the Framers found important in 1781, and what they changed their minds on by 1788.⏎They might have been onto something here:⏎Term limit for legislative office⏎Articles: No more than three out
-
remote-controlled torpedoes and devices that had applications in missile defense and computers, long after his death in 1943.⏎Now the race is on to preserve Tesla's<|endoftext|>For years, the feeding habits and nutritional requirements of the b
-
, is stranger: they became symbols.⏎There is a single species of palm native to the entire state of California, the California fan palm, which is a big one with what looks like a fuzzy beard of brown leaves underneath its green fronds.
-
Fisher (pictured above) and his band of treasure hunters finally found the big one. ‘Nuestra Señora de Atocha’ was the largest and
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
- L2:1718 suppresses
- L0:2376 weakly excites
- L2:2123 weakly excites
- L2:49 weakly suppresses
- L1:1464 weakly excites
- L2:2169 weakly excites
- L0:1981 weakly suppresses
- L1:2832 weakly suppresses
- L0:1301 weakly suppresses
- L2:111 weakly suppresses
- L2:552 weakly excites comma inside a stacked adjective list
- L0:879 weakly excites
- L2:788 weakly excites
- L2:1657 weakly excites
- L0:2320 weakly excites
- L2:960 weakly suppresses
- L2:92 weakly excites
- L2:2810 weakly suppresses
- L1:1522 weakly suppresses
- L2:1628 weakly excites
- L0:1504 weakly suppresses
- L1:2922 weakly excites
- L1:788 weakly suppresses
- L0:239 weakly suppresses numbers and dates with stray punctuation
Feeds later layers
- L5:2799 excites
- L4:146 excites
- L8:1871 weakly suppresses
- L6:2086 weakly excites
- L6:2164 weakly excites second half of multi-word proper names
- L4:1867 weakly excites
- L17:2874 weakly suppresses
- L22:844 weakly suppresses
- L15:1165 weakly excites
- L22:903 weakly suppresses
- L10:1010 weakly suppresses 'in the process of' and continuing actions
- L26:2462 weakly suppresses
- L20:1293 weakly suppresses
- L4:1106 weakly suppresses
- L18:1463 weakly suppresses
- L4:245 weakly suppresses
- L17:184 weakly excites
- L13:1035 weakly suppresses
- L14:42 weakly excites
- L4:160 weakly excites
- L23:2556 weakly suppresses
- L7:351 weakly excites
- L15:1964 weakly excites
- L8:2718 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
- 亲
- ème
- 请选择
- przy
- ыта
- éd
- avel
- .Cho
- êt
- ฉ
Suppresses
- igure
- 参与到
- opyright
- 几何
- rup
- ('('
- rush
- 같이
- wagon
- stärke