Question

Multimodal AI has a hidden problem.

Images → one tokenizer

Videos → another

3D → completely different setup

And it gets worse:

• Models that generate visuals don’t really understand them

• Models that understand visuals can’t generate them well

So instead of one intelligent system,

we end up with a stack of disconnected capabilities.

Apple is trying to take a very different approach with new model - AToken

Instead of adding more pieces, it removes them.

• One tokenizer

• One encoder

• Works across images, videos, and 3D

The core idea:

Treat all visual data in a unified format.

Images → (x, y)

Videos → (t, x, y)

3D → (x, y, z)

Everything becomes part of a single 4D token space.

So the same model can:

• Understand

• Generate

• Reconstruct

Across all formats.

And the real unlock:

Data leverage.

We have massive image datasets.

But very limited video and 3D data.

With a shared model:

→ Learning transfers across modalities

→ Less data needed overall

→ Faster capability growth

This is exactly what happened with LLMs.

One tokenizer → text, code, conversations, everything.

Now we’re seeing the same shift in vision.

From:

“different models for different media”

To:

one model that understands the visual world.