Joe Rogan Experience #2156 — Jeremie & Edouard Harris Transcript

Transcript

0:00 The Joe Rogan Experience that we co-founded. AI space? For a while in different ways so yeah we actually we started off as So he like opened up our batch at YC with this big speech it is today started then. article and a real person wrote this news article so that was an inflection make money and put that money right back into scaling up the system some more money in

3:50 but he had some connections in government, in DoD, They were saying to us that was like mundane reality like water cooler conversation at least interesting trajectory, let's say, wave of risk sets just begin to materialize. And that's kind of the weaponization risk Just as you increase the general capability like generally as smart, and we can go over these examples, we've seen example after example

7:23 we basically get disempowered in various ways So it's maybe worth, yeah, taking a step back or is something new and fundamentally different happening? That model contains, it's kind of like a human brain, increase the number of artificial neurons it contains. How does the neuron think? get stronger. The connections that are associated with doing it badly get

9:45 We have these massive systems, artificial neurons connected to each other. The strength of those to work together coherently. the amount of computing power you're gonna use to train it. is looking at this equation. in the single biggest infrastructure in human history. pretty damn soon And so when you build a data center, the sun isn't always shining, the wind isn't always blowing.

12:49 The race is you're talking about 2020 people realizing this. And as a result of the investigation we did. But we are also building better, more like so small it charts and you say, this is going to be a real problem. Like that sounds like an important problem for somebody else. we gotta figure out how to hand off our company. goal. So we start to line up as many briefings as we possibly can across the

16:07 I see the argument, makes sense. Nobody was tracking this issue. came out, the Eye of Sauron was not yet on this, and so there was a view that like about it. undo it oh yeah I mean like no matter how much money you're making the dumbest I think you know may have heard of them don't talk to the government about this. you wanna fail cheap. And we found that people are way more thoughtful

19:58 in in kind of a ready culture. But one of the issues with the individuals we to try to get the Department of Defense to build an AGI I mean that was the people who cared about the the risk. Yeah the It's all very, very open. this is part of the investigation. or no, yeah, sorry, he put his number in my phone. after the fact, where we would talk to these folks and get them in an informal setting.

23:17 It was that you had this looming cloud over everybody You go much further than that, Yes. 4-0 is close to it, or 4-0. going to be hard to control, and there's a risk to civilization So that's a different threshold and so all these different ways of defining it As the system is going to be greatest, just as the risk is greatest, a challenging thing. Yeah, that's something that one of our friends at the

27:14 So it's just a mad race to who knows what. We have a knob, a dial, we can just tune you know that that takes time the private sector is Geopolitically, economically, industrially, yeah. I mean one of the challenges is like the so so much of this is untested and we don't with that level of scale, of it before rolling it out. They don't want robots creating a billion accounts

30:54 Like, kind of calling it out. It's like, I can't tell this worker that I'm a bot because then it won't help me solve like what are the tools that we have to oversee this? And, you know, when we did our investigation, So the idea of licensing, as part of that, you need a way to evaluate systems. giant chunk of text, all of Shakespeare. you ask it, what's the best burger?

33:26 Yeah. statistical indicators that they're being tested. We've already seen them adapt their behavior to be the rule. As these systems get more scaled and sophisticated, they can pick up on more and more subtle statistical and what it potentially could put out countries, sourcing of minerals that are used in everyone's cell phones in the where if you ask it, just repeat the word company over and over

35:44 So this is called it's called rent mode internally, Now rent mode is interesting because- as far as we can tell, something around GPT-4 scale, Because it's a convergent behavior later, at least it seems to be empirically with a It's so stupid like I would never make that mistake. they're, you know, they're unintelligent in obviously very different ways.

38:38 yeah, why are all these humans making these mistakes at the same time? And so around that, yeah, that aspect, right? So there's no way to really intelligently reason. but ultimately no one knows. Humans have a very bad track record of thinking of others, other stuff as other, certainty that there was no way these systems could ever become conscious, we would still

41:52 or potentially entering an area the key theoretical lines of research in this area is is something called power No matter what my goal is, I'm probably alive, making sure that your goal doesn't get changed, accumulating power, trying to make yourself smarter. one of the risks is if you dial that up to 11 and you have an AI system that is Is it just embodying life is suffering?

45:19 It sounds speculative. It sounds wacky. The training process for these systems So imagine taking every sentence on the internet, roughly, Now, if you're gonna fill in that blank, right? you must have some understanding of how the world works. All we have is a process for training these systems. you get all this like wacky text. Okay, clearly that's indicating that somehow

47:39 it will autocomplete and give you the answer. if in fact that is what is happening. like you give the wrong answer, oh, thumbs down, you get like a little like shock or jumped for the coin great and then what they did is they moved the coin That overlapped. That's not what I meant too late. Yeah. Yeah Your guess is as good as ours. Is it saying that because it recognizes that human beings suffer and so it's taking in

51:50 I don't know what happened So it's moderated. it what you think about life, like, it's not going to really, and is it becoming a better version of Or this, let's say this model that can do cool shit. Let's train it out yeah and the again I mean this is it's a really Like it's that you know not be shut off blah blah blah There's no need for, like, I'm suffering.

55:35 You know a good starter some way of like, you know, yeah, I guess probably good idea. Yeah We're just going to do a bit of brain surgery on you, like pick out your oh, that's what you wanted, right? so you're saying like, I don't know, GDP, because people will come up with dangerously creative hacks, And if it looked like the hand put the cube on or like had correctly like grabbed the

57:56 Like opening and closing it. to give it a little bit more training game of Whac-A-Mole or whatever where you're trying to like get your values into the system, money they're investing in them and the amount of resources they're putting towards this just to break off for a second, instantaneous. There was no, no one was exactly on time for COVID. You were either too early,

1:00:56 You had to get vaccinated to fly. And then Joe's gonna be either too under caffeinated or too fast. We'll figure it out. Yeah. But So you've got all these feedback loops Because today, like today, Back then, there was no and so we looked at the landscape we talked to some of these people we investors or VCs or institutions, because you're always going to be kind of looking up kind

1:04:19 It's what you have to do because of how much money there is flowing in this space. Like is somewhat different at least but highly unusual for their level of access to the USG writ large. thought the government would kind of be you know up the creek with no no paddle and your interest in a problem, the money just comes as a side effect. The scale comes

1:07:11 absolutely could found a billion dollar company like like if I just believed my own story that follow your curiosity and interest and the Or are you likely to work at the Department of Making Sure Obviously, there's a whole bunch of challenges there. I think right now, the lay of the land is, I mean, it's looking pretty good for the US. Well, you're gonna need a lot of AI processors, right?

1:09:59 doing is actually pulling them ahead in two ways. One is when labs here in the US open source their models. at least over in China finds it extremely helpful Jeremy mentioned this but the weights are the crown jewels, right? Once you have the weights, you have the brain, you have the whole thing. are just known to be inadequate at least some of these places. And you put those together,

1:13:08 like an immediate effect of. job of blocking and tackling on the exfiltration side open sources. and there's no mechanism to prevent that from happening. on the capability side. What are they saying specifically? in the leadership team at OpenAI and OpenAI leadership were touting and that led to the departure of Jan Lejka. That's another issue. It's possible we should take them seriously.

1:16:54 again, for the second time And I think that's important. What was the stated reason why he was released? Anthropic, to OpenAI. It seems as if it's possible that Sam might have turned and a board member at the time, well, what we'll do is we'll hire you at our end. and they kind of see the writing on the wall. and it's people saying, hey, we want Sam Oltman back.

1:20:22 and that helped bring Sam Altman back. start to creep upward, there's more and more pressure more and more people interested in accelerating, to safety and whatnot, might not be the open AI of today. So it is like safety and security ultimately come out of margin, right? Some of it goes to safety you have to do an allocation of who gets what? revenue and profit gets compressed and compressed and compressed the more

1:23:32 cyber offense talent in the world. they gonna do? Drop out of the race? If they drop out of the race, Well, part of it, so part of it is that environment of deep uncertainty the question is what do you do in that it's just not going to work. It's now kind of time to get that next beat. at the frontier labs would have us treat them. And that's gonna happen soon.

1:27:13 It's a fun guessing game of what are they going to be capable of next? Mm-hmm. We need to do a better job of It's just a matter of switching from the, like, But to your point about where this is going ultimately, you know Didn't have to be that way. all the computation, all the thinking work It may make it possible for the central planner's dream to come true in some sense which then

1:30:01 is at the root of that. What happens, yeah, when you have for very centralized authorities to just be more successful. That that same thing happens except that the entity that's doing this Are we giving birth to a new life form? gives rise to consciousness because of nothing magical, because that itself is controversial, at one of these labs, It's tied into materialism because everybody wants the newest, greatest thing, so that

1:33:51 result of that behavior is not necessarily something that any individual It's because I do the thing that gets me paid Yeah. you create you know anytime you have regulation, you're going to create friction to some extent. you can't completely close off And it's really important to at least have your regulatory system allow for that possibility, This is, and it's, AI is not a special case of this, right?

1:37:33 as the thing gets proved out more and more very rapidly, Like that was really kind of shown to work If that's 2012, that's the end date of the Mayan calendar. What if that, if it is 2012, how wacky would it be were also kind of, they were floating around a little bit. Yeah, you can actually look at the like hieroglyphs or whatever and there's like neural networks. Yeah

1:40:00 I think that's actually like a completely real, I was watching this like talk by a bunch of open AI because we're actually building the thing or you're gonna just change your, And then question mark and then smiley face. That's basically the three steps that they envision Just the influence of social media on general discourse. So the wacky thing is like, the very best models now are, you know, arguably smarter

1:42:39 or they're doing whatever. I'm gonna worry about automated cyber attacks But the field is moving so crazy fast more forward-looking than you might've expected Whew, you're not making me feel better. I love all this encouraging talk, but I just, I'm playing like the leaders of any of these frontier labs That's the kind of situation that we need to be having.

1:45:34 And that's actually what excites us. And so actually confronting that head on, Not just from a technical standpoint, of that now than we ever were. Hopefully that's a good enough time horizon. And the stability of our critical infrastructure, where the money is being made and this is why these things are being implemented the and I feel fine. really you. Like that's kind of cocky.

1:49:36 where they said explicitly, We won't we won't do the thing where we just like use somebody else's voice who kind of sounds like Why would I ever take your Buddha statue, Joe? It'd be kind of hot. Everybody's a Luddite. with herself? Like, I don't know. Yeah. Yeah. I think a lot of these questions It's it's definitely a subset it's kind of desirable that it will happen and so we should plan to hand over power to AI and

1:53:01 Yeah. Whatever, but there is there's a strand of that a thread of that and a kind of like the utopian vision and I think there's a lot of that kind of driving the train right now in this space And then my question was, well, what's going to stop a guy like Donald Trump from making That we face as meat machine, whatever's like, yeah, Right. And he's like, ignorance is bliss, man.

1:55:50 It may not be human civilization if that's not what you value. That's the problem. That's the problem. Yeah one of the problems. Yes on people laying in puddles of their own jizz that's spending 8, 10 hours every day just You're running away rocket to flying over here. The things are happening. You got 3d sound this one thing that we talk about a lot in the context and Ed brought this up

1:58:38 What we're starting to see is AI optimized ads That's a really easy metric to optimize is a really good feedback metric. They click the ad they didn't click the ad But right when when you get to the point where it's like, yeah, 90% of the time, or 50 or whatever, what's that threshold that tells them, hey, you should end this. you're not allowed to do that anymore.

2:00:37 But just I'm dating a model means something different in 2024. sticking her ass at the camera. Fucking nuts. Now the virtual space is still like, pong. But it's, the gap between that and accepting it as real When that comes about, when you're getting sensory input and then you're having real and this is super speculative, It kind of, in a sense, you could argue, instantiates maybe a simulation of how the world is.

2:03:55 which could basically include like all of known physics But it kind of speaks to that question of like of consciousness to like right. What is it? Yeah There's a which allows us to have the hubris to make something like AI But when you look at like, you know cells cells communicate with their environments in ways that are completely different and alien to us

2:06:18 expressing information faster than the speed of light what do you trying to So, okay, there's... To that data, we're gonna fit some theories. Like these are the different interpretations of the theory. understanding the consciousness equation, sentience equation when it comes to like We're able to explain a whole bunch of shit. The world seems really simple

2:09:29 And the challenge is, this might actually end up being true with quantum mechanics. we'll have to refactor our whole picture of the universe You have a little monkey brain. That's just like super advanced, but it's really shitty. You know what that's harsh, but it sounded really hot So, opening eye, at one point, I don't know if it's like it. That's like, that's an optimization

2:11:31 if it does get to an AGI place, And even before AGI, because remember, theorem conjecture called the Riemann hypothesis. I'm almost tired of using the phrase of building blocks. based on its amino acid sequence, Now Alpha Fold 3 comes out, they're like, yeah, Mind and is more for by accurately predicting the structure of proteins developed by Google DeepMind in isomorphic labs

2:15:06 in the world of AI, right? that it's happening at such a strange exponential rate. I'll just tell you right now. It just never really brewed. Bullshit. This is amazing. Yeah, so there's this one paper that came out, and they're like, hey, by the way, we've increased the set of stable materials known to humanity by a factor of ten. the Middle Ages discovered.

2:17:29 that we're living through. well all the different perspectives. together. The two teams together, the State Department and us, we went to London, UK, I didn't I would not have guessed existed that was honestly speak more the stuff they made happen for us in terms of brain, of AGI and the risks there. yeah, this is a watershed moment in US history. you know, congressional hearings

2:21:06 is really important. I still don't know how to feel about Really appreciate your time. Yeah, I have this little podcast called last week in AI Aliens landing. Appreciate it.