Invest Like the Best with Patrick O'Shaughnessy
00:34:44 · Patrick O'Shaughnessy
[00:34:44] S0: One of the huge debates right now is what to do about the fact that there are open source models, not American made, which are competitive at the frontier of performance of AI models and seem to clearly have been at least to some degree based on the work of American models. What to do about this? What it means for the future of AI? Companies love open source because they can build their own thing and base 10 and others that serve a lot of inference work with a lot of these models. How are you thinking about what's right? What should happen? What will happen? The implications for business? This is a big, hard, important, interesting question.
00:35:25 · Sarah Guo, Patrick O'Shaughnessy
[00:35:25] S2: It's a big question. My point of view on what is healthy for businesses, America, the ecosystem individuals, and then there's what's already actually happened. The reality is we have over the last three years increasingly competitive open source models from all fronts, Largely China, but definitely also The US and Europe. Most recently, Thinky Poolside, people waited for reflection, NVIDIA models. Mistral, we will have very powerful and already do have very powerful open source models from Western countries. The cat is out of the bag and these are in use everywhere. Even if you have no economic point of view on the labs and you just say, how is the diffusion of capability going to happen in the economy? There are a huge number of instances where it is too expensive, too sensitive, or too slow to use the model from the frontier providers today. And I think that's going to increase as we learn how to do more things with AI because it's actually quite expensive. It's objectively true that if the capabilities are more democratized, you will see them used in more ways. I want to see that happen. I do think that it would be irresponsible not to understand the safety profile of models as they progress because you just draw the line. We have companies that use frontier model capability for defensive cybersecurity and for biology work. If it works for those use cases, it obviously also should work in similar ways for [00:37:07] S0: the office of those
00:37:08 · Sarah Guo
[00:37:08] S2: use cases or the bioweapons, biosecurity use cases. We just need to look at that reality and think about what are the other ways in which you control this. But attempting to stop technological progress and openness around it. If you did restrict use of open source models in The United States, you'd basically just restrict law abiding American businesses and slow them down or move profits, different pockets, prevent certain uses of them because the actual attackers or people who have adversarial uses of these things are not affected by your restrictions. You're restricting your own people. My view would be there should be testing and understanding of these models at the frontier. People are very worried about backdoor like behaviors in Chinese models. The thing to do would actually be, like, have a very rigorous set of safety testing on that. Like, let's go find out as much as we can instead of talking about how there might be this issue in a speculative way where there's not been nearly enough actual research on it. The future where there is broad access to intelligence too cheap to meter, as Sam put it, that is coming. It will be supported by open source. Businesses want it to control their own destiny for economics, for capacity. If given those models and the increasing democratization of the skills to post train these models, build harnesses, use tasks, The economy is so big. Every individual has these use cases that are not going to be imagined by a researcher in a frontier lab. You can't imagine the diversity of reality. Even if you trusted the models to go figure out what to do, they have to get there. The best way for that to happen is a ecosystem of businesses as we've always had in the economy, cheap infrastructure.
00:39:07 · Patrick O'Shaughnessy, Sarah Guo
[00:39:07] S0: Do you ever worry that an alternate version of US history is that there was energy too cheap to meter because we built a thousand AP one thousands or something, much like China is doing now in nuclear. And just a set of circumstances happened such that, like, we just didn't get that. Do you ever worry about that as it relates to intelligence? It does seem inevitable that we're gonna have abundant, accessible, low cost, valuable intelligence. Can you imagine a world where we don't? [00:39:31] S2: Yes. Absolutely. And I also think I can imagine very easily a world where we don't have that in a competitive way because it is essential to economic competitiveness and national security. There is not a version of the world where we rebuild our industrial base without automation in The United States. If we don't import people and our people are expensive and we lack some of the skills, want to produce a lot more goods and have a more resilient supply chain. [00:40:00] S0: Just doesn't add up.
00:40:00 · Patrick O'Shaughnessy, Sarah Guo
[00:40:00] S0: Just doesn't add up. [00:40:00] S2: Who's gonna produce this stuff? People in The United States do not want to work and should not want to work for $13 an hour doing a very inhuman job. Don't I think it is inevitable that we are competitive. And I think we need to like make that decision actively. The version of it that I think is very possible is that people are rationally afraid of the impact of AI on jobs or dislike the capture of rent by a small number of technology firms, rejecting the idea of being in the permanent under class and then connecting that to a anti capitalist orientation. That contingent of thought can slow down build of energy and infrastructure and industrial capacity. One of the most important inputs is compute. If we don't have it, we're naturally not competitive or we're at least not independent. I think we're gonna start talking much more about compute independence. That's a big problem.
00:41:00 · Patrick O'Shaughnessy
[00:41:00] S0: So on this point of compute independence, what does that mean? What are the missing pieces of compute independence? Like, the most obvious one might be more fab... More leading edge fab capacity here in The US or something like this. There's all sorts of stuff upstream. There's particular kinds of glass, very important that's controlled by basically one company that TSMC has like monopoly on the supply of. There's all these component parts. But if you think about compute independence, like if so much boils down to compute, what's to be done about it? Are you trying to invest in companies that are solving that problem? What is the problem? Say a bit more about what it'll take.