"Patrick, over to you. All right. Thank you, Graeme. Um, and, uh, and thank you, Sam, uh, for being with us. Uh, last year, I actually interviewed Sam Bankman Freed, which was, uh, which was clearly th..."

"Uh, so, um, so we we'll start out with the topic on, uh, on everyone's mind. Uh, so, uh, when will we all get our world coin? Uh, I think if you're not in the US, you can get one in a few weeks. If yo..."

"ever. I don't know. All right. Um, so, uh, which is a crazy thing to think about that like this is, you know, think whatever you want about crypto and the ups and downs, but the fact that the US is th..."

"know, testing something just you actually want to get you where chat GBT is purely, you know, an instrumental tool for you? Summarization by far. Uh I I've gotten like I don't know how I would still k..."

"possibility. If we end up in the world where we asmtote soon, what do you think kind of exposed we will, you know, look back on the reason as having been too little data, not enough compute, what's wh..."

"we're we're now training on kind of order of all of the internet's tokens and you can't grow that, you know, another two orders of magnitude. Uh I guess you could counter with yeah the synthetic data ..."

"the plan. Mhm. If um so one of the big breakthroughs in I guess uh GPD 3.5 and four is RHF. Uh um you know if you Sam uh personally sat down and did oral all of the RLHF would the model be significant..."

"And h how many how many like how should one think about the question of how many smart grad students one needs like is one enough or do you need like 10,000? It's we're studying this right now. We we ..."

"of required the power of nations and we made the IAEA which I think was a good decision on the whole and a whole bunch of other things too. So like yeah, I think probably anything you can do there to ..."

"AI supercomputers do have some similarities and this is a place where we can draw more than usual parallels and inspiration. But I would caution people to to overlearn the lessons of the last thing. U..."

"this would be a capabilities threshold, but that's harder to measure. any any system cape over that threshold I think should submit to audits um full visibility to that organization be required to pas..."

"So one of the there's also I think there there there's like a bunch of unusual things about this is why it's dangerous to learn from any technological analogy of the past. There's a bunch of unusual t..."

"closed source models and there will be the progress that the open source community makes and it'll be you know a few years behind or whatever a couple years behind maybe um but I think we're going to ..."

"many super economically valuable things, yes, the smaller open source model will be sufficient. But you actually just touched on the one thing I would say, which is like help us invent super intellige..."

"Um yeah I mean a lot like I think that's most of the new work between here and super intelligence will move that probability up or down. Okay. Is there anything you're particularly paying attention to..."

"And do you think sufficient interpretability work is happening? No. Um why not? You know, a lot of people say they're very worried about AI safety. So that seems, you know, superficially surprising. M..."

"single people or small groups of people that are very technical that want to push forward a technical solution. um and are you know maybe in grad school or just out or in undergrad or whatever I I thi..."

"program at OpenAI that does exactly this, and I'm astonished how well it works. It seems that pretty soon we'll have um uh agents that you can converse with in very natural form, low latency, full dup..."

"is important is that we we we establish a societal norm soon that you know if you're talking to an AI or a human or a sort of like weird AI assisted human situation. Um but people people seem to have ..."

"of AIs integrated along with humans and yeah there have been movies about this for like a long time like there's like you know C3PO or whatever you want in Star Wars like people know it's a AI. It's s..."

"I I would take the other side of that. Again, we don't know. But I I would say that like human cap capital allocation is so horrible that if we know exactly what to do, even if it's expensive, you you..."

"um and building the org that can make these repeated breakthroughs. Uh they don't all work. You know, we like went we've gone down some bad paths, but we have figured out more than our fair share of t..."

"starts tomorrow. Um, if you were CEO of Google, how would you do? I think Google's doing a good job. Um, I think they they they have had like quite a lot of focus and intensity recently and are really..."

"Um, how much important ML research comes out of China? I would love Sorry, go ahead. I would love to know the answer to that question. Like how much does it come out of China that we get to see? Not v..."

"I just feel confused. Um, would you prefer OpenAI to um to, you know, figure out a 10x improvement to training efficiency or to inference efficiency? It's a good question. Um it sort of depends on how..."

"there a GPT2 moment happening now um there's a lot of things we're working on that I think will be GPT2 like moments uh if they come together but nothing there's nothing like released that I could poi..."

"some kind of Siri plus sort of thing. Yeah. Yeah. And you mentioned you know curing cancer. uh is there an obvious application of these techniques and technologies to science that again you think we h..."

"same a similar system could go off and start to read all of the literature, think of new ideas, do some limited tests in simulation, email a scientist and say, "Hey, can you run this for me in the wet..."

"you know currently easy to work with I I really don't know. I I this is like, you know, most areas I I am willing to like kind of give some rough opinion. In this one, I never have un I don't I don't ..."

"more work. Um uh you know, OpenAI has a um has done a super impressive job of fundraising and has a very unusual capital structure uh for you know the nonprofit and the Microsoft deal and like all all..."

"Um, and I understand why. Like it's also great for companies that like only ever have to raise a few hundred thousand or a million dollars and get to profitability. But I think we over pivoted in that..."

"never met another person that I think I can that can be developed easily into another Elon. He is sort of this like strange N of one character. Um I'm happy he exists in the world of course but you kn..."

"the you know there's a finite or essentially finite uh set of um of funding models in the world uh And each has a particular set of incentives and for the most part a particular sociology uh and you k..."

"this. You still need like the Elon like people to to do it. Um, and like one project I've always been tempted to do, um, is say, "Okay, we're going to identify the, let's say, 100 most talented people..."

"That's kind of the university model, I guess. Uh and I don't mean that as like a this already exists. You know, you're just, you know, reinventing the bus or something. H I mean that like it's it's ma..."

"time, worked with for a long time. I think that's really valuable. Like in the case of OpenAI, uh I had known Greg Brockman for a long time. I met Ilia for maybe only like a year before, even a little..."

"have been taking it more seriously than others? Um what do you think the likelihood is that we will come to realize that GPT4 is somehow significantly overfit on the problems uh you know in the domain..."

"thinking all this AI safety stuff how if at all do you think synthetic biology should be regulated? I mean, I would like to not have another synthetic path to gym cause a global pandemic. I think we c..."

"Um and so I think the particular thing like if it is true that uh co was engineered uh I think you know instances of that set of slight modifications to already existing uh infectious diseases we can ..."

"Yeah, I I I I very much agree with that. and and and clinical trials um uh like that that was the you know the the limiting step uh in uh in co and I I think it's you know it's at this point been wide..."

"different domains um more more kind of the Henry Adams curve realized uh and they frequently observe that permitting in the broadest sense all sorts of you know well-intentioned but self-imposed restr..."

"used to or believe less. when we first met whatever it was uh 15 or so years ago, uh Mark Zuckerberg was preeminent in the technology industry and in his 20s and you know not that long before then uh ..."