It starts with a single number.
…a dimmer knob on a wire. A signal flows in along the wire. The knob decides how much of it gets through — full blast, a trickle, or flipped backwards.
Here is one real-ish weight: −0.4231
Now the technical name, since you've felt it first: engineers call this number a weight (or parameter). The network learns millions of these knob settings during training.
Playground · turn the knobs
Negative weight? The signal comes out flipped — like the knob is wired backwards. That's not broken; networks use this constantly.
…that a weight is a stored fact — "the AI memorized this." Actually one weight means nothing on its own. It's a knob setting, not a fact. Meaning only appears when millions of knobs work together.
One weight is a dimmer knob: it decides how much of a signal passes through — and nothing more.
Dozens of knobs, one speaker.
…a mixing board. Four faders feed one speaker. Each fader has its own knob setting (weight). The speaker plays the mix — then squishes it so the total can never blow out the speakers, no matter how loud the inputs get.
The jargon, now that you've got the picture: each input is multiplied by its weight, everything is added up (the weighted sum), then passed through a nonlinearity — the "squish." That whole assembly is one neuron. Real ones have dozens to thousands of inputs.
Playground · mix the board toy neuron
The squish here is tanh — it squeezes any sum, however huge, into a tidy −100%…+100%. Try maxing every fader: the output refuses to blow past 100%.
…the neuron is "thinking," like a tiny brain cell deciding things. Actually it just multiplies, adds, and squishes. Nothing in this section is intelligent — the intelligence only shows up when millions of these do it together.
A neuron multiplies, adds, and squishes — the whole trick, repeated billions of times.
Now stop counting individuals.
…a stadium crowd holding colored cards — blue for negative, red for positive, brighter for bigger. From the stands you see the pattern. Up close, each card is still just one number.
Every square below is one weight — 100 × 100 = 10,000 of them. That's roughly one small layer of a real network. Simplified illustration
Playground · one small layer
You can still almost see individuals here. Remember this feeling — it's the last time you'll have it.
…10,000 numbers sounds like a lot for AI. Actually it's barely one layer. A single layer in a big model can hold millions of weights — this whole grid would be one pixel of it.
At 10,000 you can still almost point at individuals — the last scale where a weight is a thing you can see.
Individuals are gone. Only texture remains.
…grains of sand. One grain means nothing. A beach isn't "many grains" the way a pile of coins is — it's a texture. A billion weights is a beach.
Below: 1,000 dots. Each dot is one million weights. Together they are one billion — 109. Simplified illustration
Playground · count the billion
…"a billion" is just a bigger pile — something you could sort through given time. Actually no human timescale works. Count one weight per second and you finish in 31 years. At this scale there are no individuals left to meet.
Count one weight per second and you'll finish in 31 years. A billion isn't a pile — it's a texture.
Meet Lyra 2.0.
This isn't a toy. Lyra 2.0 is NVIDIA's open-source model that turns a single image into an explorable 3D world. Today we measured its actual weight files on Hugging Face — 28.7 GB of them. At 2 bytes per weight, that's roughly 13–15 billion knobs.
They don't arrive as one blob. They come in five boxes, each with a job. Tap each box:
Five boxes of knobs · ≈14.4B weights
Now the dot field. Each dot = 100 million weights. That's the same beach as before, fourteen times over:
Playground · ~145 dots of Lyra
| Component | Weights file | ≈ weights |
|---|---|---|
| Text encoder | 11.4 GB | 5.7B |
| 3D reconstruction | 13.4 GB | 6.7B |
| Image encoder | 2.4 GB | 1.2B |
| VAE | 0.5 GB | 0.25B |
| 3 LoRA adapters | 1.1 GB | 0.55B |
| Total | 28.7 GB | ≈ 13–15B |
How we estimated: the weight files store 16-bit floats — 2 bytes each — so gigabytes ÷ 2 ≈ billions of weights. Measured from the real files at huggingface.co/nvidia/Lyra-2.0, October 2026. It needs an H100 with 80 GB of memory just to run.
…28 GB means 28 GB of stored 3D worlds. Actually it's 28 GB of knobs. No world, no image, no video is stored anywhere in there — you'll see why in the next section.
Lyra 2.0 is ~14 billion dimmer knobs in five boxes — and not one of them is a picture of anything.
The weights are not a photo album.
Here's what almost everyone gets wrong: none of those 28 gigabytes "looks like" anything. There is no tiny 3D world folded up inside. No hidden pictures. Just numbers.
…a recipe vs. a cake. The weights are the recipe — 14 billion tiny instructions like "a pinch more of this." The 3D world is the cake: it doesn't exist until you bake it, fresh, every time you run the model. You can't find the cake inside the recipe.
Playground · peek inside the 28 GB
Real-format values, drawn from the same kind of distribution as trained weights. toy values
When Lyra turns your photo into a walkable 3D world, it reconstructs it at runtime — signals flowing through 14 billion knobs, each nudging the result a hair. Nothing is retrieved. Nothing was ever stored.
…AI "remembers" things the way a hard drive does — that somewhere in the weights sits a copy of every image it trained on. Actually training smeared everything across billions of knobs. Ask for one training image back and you can't point to where it lives — because it doesn't live anywhere.
A neural net doesn't contain worlds — it contains the knobs to rebuild one, fresh, every single time.