2 ms·

Think of all the books that will be watched when we have text-to-film models. Thousands of years of content to be depicted in a number of ways.

by _akhe 2y ago

tazu 2y ago
The first thing I would love to see is Borges. Specifically, his Library of Babel. I love listening to Borges audiobooks while on heavy doses of mushrooms, so I have a pretty vivid visualization of all of them already, but it would be great to see them so I could point at the screen and be like "whoa, yup, that's it!"
- _akhe 2y ago
  You may end up being a great film director in the near AI future thanks to your studies in psychedelia :) I'm most excited to see what people who would otherwise never make a film might come up with
jiggawatts 2y ago
The distant future of generative AI is essentially a Star Trek holodeck. As an intermediate step I can imagine a future VR headset that coupled to an AI accelerator that can output 8K per eye in real time based on the context of an entire novel, TV series, or collected extended universe works of fiction.
- _akhe 2y ago
  You just made me realize how slow our LLMs are today - we're like in the vintage years! That's an incredible vision... and not too far away.
  - jiggawatts 2y ago
    In another thread I likened the current era of AI to playing with PovRay in the 1990s to render "Amazing raytraced 3D graphics!" at a snails pace, with pixels crawling across the screen left-to-right, top-to-bottom over hours or days. Some of use got very excited even back then, because it gave us a peek at a likely future or real-time photo-realistic 3D graphics. We're there now, a mere three decades later. I expect the pace of AI research to be vaguely similar. In three decades, we'll have interactive generative VR for sure. In the meantime, we'll hit a lot of fun milestones too. I suspect the first games that are real-time AI generated are not that far off. It's surprisingly easy to train AIs when there's simulation available as the ground truth, especially when that simulation can be modified to output per-pixel embeddings, not just noisy RGB sensor measurements like with video input training data.

Think of all the books that will be watched when we have text-to-film models. Thousands of years of content to be depicted in a number of ways.

by _akhe 2y ago

Think of all the books that will be watched when we have text-to-film models. Thousands of years of content to be depicted in a number of ways.

tazu 2y ago

The first thing I would love to see is Borges. Specifically, his Library of Babel. I love listening to Borges audiobooks while on heavy doses of mushrooms, so I have a pretty vivid visualization of all of them already, but it would be great to see them so I could point at the screen and be like "whoa, yup, that's it!"

_akhe 2y ago
You may end up being a great film director in the near AI future thanks to your studies in psychedelia :) I'm most excited to see what people who would otherwise never make a film might come up with

jiggawatts 2y ago

The distant future of generative AI is essentially a Star Trek holodeck. As an intermediate step I can imagine a future VR headset that coupled to an AI accelerator that can output 8K per eye in real time based on the context of an entire novel, TV series, or collected extended universe works of fiction.

_akhe 2y ago
You just made me realize how slow our LLMs are today - we're like in the vintage years! That's an incredible vision... and not too far away.
- jiggawatts 2y ago
  In another thread I likened the current era of AI to playing with PovRay in the 1990s to render "Amazing raytraced 3D graphics!" at a snails pace, with pixels crawling across the screen left-to-right, top-to-bottom over hours or days. Some of use got very excited even back then, because it gave us a peek at a likely future or real-time photo-realistic 3D graphics. We're there now, a mere three decades later. I expect the pace of AI research to be vaguely similar. In three decades, we'll have interactive generative VR for sure. In the meantime, we'll hit a lot of fun milestones too. I suspect the first games that are real-time AI generated are not that far off. It's surprisingly easy to train AIs when there's simulation available as the ground truth, especially when that simulation can be modified to output per-pixel embeddings, not just noisy RGB sensor measurements like with video input training data.