large number scale words
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. 4 of Qwen3-0.6B. Its strongest snippets are fairly consistent: several of them 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 responds to big number words like 'billion', 'trillion' and 'bn' that come right after a figure. When it was turned up, the model wrote out large numbers and population counts.
Written by Claude from this neuron's activations and steering outputs. Confidence: high. 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.
-
orbiting satellites collecting solar energy and blasting it down to receiving stations in high powered laser beams. Safe free energy forever .... all we need are a few hundred thousand billions to set it up.<|endoftext|>Built on our successful pilot work in two Bay Area
-
internet becomes integrated into virtually every aspect of life, keeping its infrastructure secure will be ever more essential. Spending on cybersecurity globally is expected to surpass $200bn in 2
-
The development of hydraulic fracturing (‘fracking’) has raised US gas production dramatically, and helped double the estimated global gas resource to around 800 trillion cubic meters (tcm), or 250 years’ consumption at current
-
of processors. The country had crossed the petaflop computer range and needs exascale computers, which are machines capable of performing 1000 trillion operations a second."⏎IIT Delhi will provide the infrastructure for the Exascale
-
in about a cup of DNA.⏎There is a lot of digital information in the world about three zettabytes' worth (that's 3000 billion billion bytes) and the constant influx of new digital content poses a real challenge for
-
the power system by 2030, as much as 15 trillion litres of water could be saved. That's the equivalent to the annual consumption of
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: I am doing, something that is I am doing, and so I am (something) This is the same as I am (something) This is the same as I am This is the same as I am This is the
Most negative on
Where the activation went furthest below zero. Blue marks negative activation. Many neurons mean something different in each direction.
-
of the neuron network of a human brain, with ultra-high resolution and in three dimensions. The coalition now grows to a network involves over 1,000 researchers from Singapore, Australia, China, Japan, Korea and Taiwan. The team
-
as "the Monster" to the scientists who conjured it – to explore the mysterious subatomic realm. Fifty years later, more than 1,000 people gathered at SLAC National Accelerator Laboratory on Friday and Saturday to celebrate the
-
dust generation inside pebble-bed reactors.⏎MELCOR is a fully integrated, engineering-level computer code that models the progression of severe accidents in light water reactor nuclear power plants. MELCOR is being developed at Sandia National Laboratories (
-
023/B:BIOC.0000048451.07820.9c⏎- 1.1k Downloads⏎Recent massive imports of slider turtles, (Trachemys scripta
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
- L3:2436 suppresses
- L2:1987 excites
- L3:1499 weakly suppresses
- L3:1510 weakly suppresses
- L1:1464 weakly excites
- L3:294 weakly excites
- L1:1099 weakly excites
- L3:2671 weakly suppresses
- L3:189 weakly excites
- L3:359 weakly suppresses
- L1:951 weakly excites
- L3:829 weakly suppresses
- L3:1224 weakly suppresses
- L3:1827 weakly suppresses
- L2:3033 weakly excites
- L1:1973 weakly suppresses
- L1:146 weakly excites
- L1:2846 weakly excites
- L3:2140 weakly suppresses
- L2:483 weakly suppresses
- L3:2852 weakly excites
- L1:1177 weakly excites
- L1:1229 weakly excites
- L3:115 weakly suppresses
Feeds later layers
- L8:787 excites
- L5:1662 excites
- L7:1085 excites last digits of big numbers
- L6:560 weakly excites
- L8:1103 weakly excites
- L8:2397 weakly excites
- L10:2406 weakly suppresses
- L18:1259 weakly suppresses
- L15:2476 weakly suppresses
- L8:620 weakly suppresses
- L13:2568 weakly excites
- L24:2545 weakly excites
- L8:280 weakly excites
- L21:1382 weakly excites
- L25:3070 weakly suppresses
- L19:2899 weakly excites
- L11:1820 weakly suppresses
- L6:1099 weakly excites
- L9:2006 weakly excites
- L16:1606 weakly excites
- L23:297 weakly excites
- L11:1304 weakly suppresses
- L23:2696 weakly suppresses
- L11:1215 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
- ordo
- acre
- básica
- ollider
- rigs
- estyle
- fetch
- 氪
- ffects
- Sark
Suppresses
- 峃
- 写作
- abs
- 绳
- integrating
- ButtonModule
- ichever
- integration
- :index
- writing