Reasons AI world models don't "scale":
- 3DGS is not efficient for rendering nor download size
- 3DGS does not support dynamic lighting nor physics nor nav meshes
- AI world models don't support "offline rendering" (eg streaming tiles)
- to make an AI world model of Earth, do we just take photos of every square inch of the planet and use that to describe our planet?
- no mechanism for audio
- not deterministic so not good to achieve simulation training outcomes
- and I'm sure other things cuz there's so much that goes into building a virtual world
Re. space data centers, I haven't given it much thought, nor crunched any numbers, but we rely on undersea fiber optic cables for the intercontinental backbone of the internet. The latency in sat comms is too high.
So we'll build *slow* data centers in space...?
a16z's Big Ideas 2026 includes two sections on AI world models being interactive, gameable, etc.
What a crock of shit. Not one mention of how these worlds will render at high framerates, be playable across wide areas, and many other things game devs do today to make open virtual worlds.
The AI media bow wave is close to creating a vacuum back to the reality of building large-scale open worlds.
Why the fuck isn't someone leveraging Mapillary's open, street-level imagery to generate 3D content around the world?!
I've got a ton of ideas to make a shit-ton of money. All I need is a rendering guru to join me and the world will be our oyster.
Hook a guy up, peeps!
Unique perspective, love it!
There's a slight catch 22 here in that, in order to hire these high performers, the founders must've convinced investors to invest so they could pay the early engineers. That implies the investors saw the opportunity first, or at least before the early employees.
Regardless, the points made in the article still stand. Thx!
@lucyhargreaves4 Wow. Ripe for massive lawsuits given the generally-internationally-recognized, and de facto, standard that the presiding law is that of the country where the servers sit!