In conversation with Sam Altman

At some future point (no specific date given), AI systems with a high level of intelligence will have latency that feels instantaneous to human users and will be so inexpensive to run that their marginal usage cost is effectively 'too cheap to meter.'

“we want to like, cut the latency super dramatically. We want to cut the cost really, really dramatically. Um, and I believe that will happen. We're still so early in the development of the science and understanding how this works. Plus we have all the engineering tailwinds. So I. I don't know. Like when we get to intelligence too cheap to meter and so fast that it feels instantaneous to us and everything else. But I do believe we can get there for, you know, a pretty high level of intelligence.”

Mainstream phone manufacturers will, in the future (no specific date given), ship phones with specialized on-device chips designed specifically to efficiently run phone-sized AI models.

“Then would there be like a specialized chip that would run on the phone that was really good at powering a, you know, a phone size AI model? B: Probably. But the phone manufacturers are going to do that for sure.”

OpenAI will, in the future (no specific date given), significantly improve the responsiveness and naturalness of ChatGPT’s voice mode so that it is noticeably smoother, faster, and more organic-feeling than at the time of this interview (May 2024).

“We are working on that. It's it's so clunky right now. It's slow. It's like kind of it doesn't feel very smooth or authentic or organic, like we'll get all that to be much better.”

Within a relatively near-term horizon (implied to be less than multiple decades from May 2024), frontier AI systems will become capable of causing significant global harm if misused or uncontrolled.

“But I think there will come a time in the not so distant future. Like we're not talking like decades and decades from now, where AI says the frontier AI systems are capable of causing significant global harm.”

Throughout the deployment lifecycle of GPT-4–level models, they will not be the direct cause of a global catastrophic risk event (e.g., mass loss of human life or civilization-ending scenario), although smaller-scale problems and incidents are expected.

“At the current strength of models? Definitely. Some things are going to go wrong and I don't want to make light of those or not take those seriously. But I'm not like I don't have any like catastrophic risk worries with a GPT four level model.”

GPT-4–level models will not, in practice, pose a material threat of enabling recursive self-improvement or autonomous design and deployment of bioweapons (or similarly extreme threats), although minor or non-material misuse risks may exist.

“you know, we should have safety testing on the outputs at an international level for models that, you know, have a reasonable chance of, of posing a threat there. Uh, I don't think like GPT four, of course, does not pose any sort of. Well, I won't say any sort because we don't I don't think that GPT four poses a material threat on those kinds of things.”

The YC Research universal basic income study that began roughly five years before this interview (around 2019) will publicly release its results in the near future after May 2024 ("very soon").

“Our results on that come out very soon. Just it was a five year study that wrapped up, um, or started five years ago.”

Artificial general intelligence (AGI) will inevitably be created and, net, will be tremendously beneficial to humanity, rather than primarily harmful.

“I think it is unavoidable that this is going to happen. I also think it's going to be tremendously beneficial.”

OpenAI will become and remain one of roughly four dominant, most important companies in the AI sector over the coming years.

“I think that these guys are going to be one of the four major companies that matter in this whole space.”

Over time, frontier AI foundation models from leading providers will converge to roughly similar, very high capability levels, and most of the economic value and monetization will shift to the surrounding "scaffolding" (infrastructure, tools, and application layers) rather than the core models themselves.

“the models will roughly all be the same, but there's going to be a lot of scaffolding around these models that actually allow you to build these apps... I think the right way to think about this now is the models will basically be all really good. And then it's all this other stuff that you'll have to pay for.”

Companies that successfully build the non-model "scaffolding" (infrastructure, tools, and application layer) around large language models will be able to build very successful, large businesses in the AI era.

“Whoever builds all this other stuff is going to be in a position to build a really good business.”

Apple will primarily use its very large excess cash to repurchase its own stock rather than to make large new investments or acquisitions, continuing a heavy buyback strategy in the coming years.

“It's an incredible business that has so much money with nothing to do. They're probably just going to buy back the stock. Just a total waste.”

AlphaFold 3–style AI systems will usher in a new era of drug development and biological design, leading to profound, large-scale impacts on human health and the biopharma industry over the next couple of decades.

“it's going to usher in a new era of drug development, design for human health. So all in all, I'd say it's a pretty astounding day... the capabilities and the impact are going to be like, I don't know, I know I say this sort of stuff a lot, but it's going to be pretty profound.”

In the future, AI systems like AlphaFold 3 will enable in-silico search over chemical space to design combinations of molecules that can safely reprogram and "de-age" human cells (Yamanaka-factor-style approaches), leading to powerful longevity or rejuvenation therapies.

“We can now simulate that. So with this system, one of the things that this AlphaFold three can do is predict what molecules will bind and promote certain sequences of DNA, which is exactly what we try and do with the Yamanaka factor based expression systems and find ones that won't trigger off target expression. So meaning we can now go through the search space in software of creating a combination of molecules that theoretically could unlock this fountain of youth to de-age all the cells in the body and introduce an extraordinary kind of health benefit.”