00:23:08 · Jordi Hays, John Coogan, Luca Ferrari
[00:23:08] S1: The company is focusing on AI that will help in engineering and manufacturing in a number of fields, including computers, aerospace, and automobiles. Mhmm. Uh-huh. Unclear where the company will be based. Bezos' cofounder and co chief executive is Vik Bazaraj, a physicist. [00:23:25] S0: AI that helps in manufacturing of computers, aerospace, and automobiles. So computers, he manufactures racks at AWS. Aerospace, he manufactures rockets at Blue Origin. Automobiles, he's a big backer of Rivian. He manufactures cars there. And so he wants to do some sort of automated, like, supply chain. Is it ERP? [00:23:52] S1: I mean, this is clearly not PR that they wanted to do. Yeah. So I think there's a lot of guessing going on. [00:23:57] S5: Okay. Well. [00:23:59] S1: But Mr. Bezos' co founder and co chief executive is Vic Bajaj, a physicist and chemist who work closely with Google's co founder Sergey Brin. Sergey Brin at Google's X, a research effort often called the Moonshot Factory. Google X produced a wide range of ambitious projects, including Wing, a drone delivery service, and the self driving car that became Waymo.
00:23:59 · Jordi Hays, John Coogan, John Tennant
[00:23:59] S1: But Mr. Bezos' co founder and co chief executive is Vic Bajaj, a physicist and chemist who work closely with Google's co founder Sergey Brin. Sergey Brin at Google's X, a research effort often called the Moonshot Factory. Google X produced a wide range of ambitious projects, including Wing, a drone delivery service, and the self driving car that became Waymo. [00:24:22] S0: It's so interesting how divergent those two paths were. Like Wing, you don't hear about very much. When you think drone delivery service, you think Keller at ZipLine. They're the ones that are really running away with that compounding. I don't know the status of Wing. Maybe I'm just out of the loop on that. Maybe it's doing great. But it feels like Wing has not certainly garnered the level of attention that Waymo did, which was this success out of the exact same sort of incubator. So interesting. Hudson [00:24:55] S1: comments in the chat says, it's just gonna be robots doing science on the material level, not Materials. Scale. Interesting. Materials. There's more details here. Project Prometheus is among a wave of companies focused on applying AI to physical tasks, including robotics, drug design, and scientific discovery. Mhmm. Last year, Bezos invested in physical intelligence, a startup that is applying AI to robots. [00:25:19] S0: Okay. So And building okay. [00:25:24] S1: Anyways, we'll have to get Jeff on Yeah. When he's ready to talk [00:25:28] S0: about where did the money come from exactly? Do we know? [00:25:31] S1: I think it was [00:25:31] S4: Just Bezos and stuff.
00:25:31 · John Tennant, Jordi Hays, John Coogan, Tyler
[00:25:31] S4: Just Bezos and stuff. [00:25:32] S1: Mostly Bezos. [00:25:35] S0: It's interesting. It's like, yeah. But like, I I I'm I'm super interested in like how you size a round if you're investing in your own project. Because you could just be like, this is my thing. I'm gonna I'm gonna fund it every payroll cycle. Yeah. Whatever the bill is, I'll pay it because it's my thing. You don't necessarily need to, like, do some sort of funding round necessarily. I don't know. [00:25:59] S1: I Why you laughing? Accidentally opened the the comment section of the New York Times article that that we were just going through [00:26:06] S0: and [00:26:06] S1: it says, first comment, these large ego models seem promising. Large ego models? Like, if you're if you're a, you know, a tech billionaire Yeah. Yeah. And you have a large ego, you you wanna you want your own large language model. [00:26:21] S0: Yeah. It sounds like he's not very much not training just another LLM. Yeah. But we'll have to see. Well, Jan Lakoun was in The Wall Street Journal's weekend paper, profiled by Megan Barrowski. An AI pioneer thinks everyone wrong everyone is wrong again. He's been right about AI for forty years now. He thinks everyone's wrong. What do you think, Tyler? Do you think Ian Lecun is wrong or do you think everyone else is wrong? [00:26:51] S2: I mean, so in this article, he doesn't there's nothing really new here. They're just kind of talking about his the points that he's been making over the past couple years, which is just that, like, LLMs will not bring us to AGI or Yeah. Or anything.
00:26:51 · Tyler, John Coogan, Luca Ferrari, Jordi Hays
[00:26:51] S2: I mean, so in this article, he doesn't there's nothing really new here. They're just kind of talking about his the points that he's been making over the past couple years, which is just that, like, LLMs will not bring us to AGI or Yeah. Or anything. [00:27:03] S0: And Yeah. [00:27:03] S2: Like, even if you keep scaling, they don't they're not, like, actually intelligent. They can't reason or whatever. [00:27:09] S0: Yep. [00:27:10] S2: Which I don't know. It's like, these models are much better than me at math. Like, they can do they can get IMO gold medals [00:27:17] S0: Yes. [00:27:17] S2: Which, like, I cannot do. And that's like it it's like, does that take reasoning? Never doubt yourself. [00:27:21] S0: But, I mean, like yeah. The the I mean, the the there is an element of, like the computer has been able to do good math fast forever, Like, since, like, the eighties. Like, if you were, like, what is 7642 Yeah. I mean, I I think there's difference [00:27:37] S2: you can you can make between just, like, raw calculation and, like, how to, like, think about solving a math question generally. And, like, you can say that he's been right about, like you could say that he predicted, like, spiky intelligence maybe of models. [00:27:50] S5: Yeah. Yeah. [00:27:51] S2: Yeah. And that's, like, sure. Yeah. [00:27:53] S0: Yeah. [00:27:54] S2: But I I I yeah. It's like I I I I think it's not true to say that he's been right about AI for the past forty years. [00:28:00] S1: Oh. That seems man's never had a bad take. Zone. Forty years. Not a single bad take.
00:28:00 · Jordi Hays, John Coogan
[00:28:00] S1: Oh. That seems man's never had a bad take. Zone. Forty years. Not a single bad take. [00:28:05] S0: Single bad take. Now he thinks everyone's wrong. And, yeah, it's very it's very funny for me because oh, does he think George Hotts was wrong when George Hotts said that GPT six will not be AGI and the GPT paradigm will not scale on the Lex Friedman podcast in 2021? Does he disagree with that too? Does he disagree with Andre Carpathi saying it's sloth and that we need new ideas? Or, you know, there's, like, seven other people that have kind of echoed the same thing. He doesn't think everyone's wrong. He just thinks some people are wrong. [00:28:35] S1: But Sort of. I mean, Tyler's take I don't know if now is the right time I can go through this Karpathy post. [00:28:41] S0: Oh, yeah. Please.
00:28:43 · Jordi Hays
[00:28:43] S1: Karpathy posted yesterday. He said sharing an interesting recent conversation on AI's impact on the economy. AI has been compared to various historical precedents, electricity, industrial revolution, etcetera. I think the strongest analogy is that a is that of AI as a new computing paradigm, software two point o, because both are fundamentally about the automation of digital information processing. If you were to forecast the impact of computing on the job market in the nineteen eighties, the most predictive feature of a task job you'd look at is to what extent the algorithm of it is fixed. I e, are you just mechanically transforming information according to rote, easy to specify rules, I e, typing, bookkeeping, human calculators. Back then, this was the class of programs that the computing capability of that era allowed us to write. With AI now, we are able to write new programs that we could never hope to write by hand before. We do it by specifying objectives, I. E, classification accuracy or reward functions. And we search the program space via gradient descent to find neural networks that work well against that objective. This is my Software two point zero blog post from a while ago. In this new programming paradigm then, the most predictive feature to look at is verifiability. If a task job is verifiable, then it is optimizable directly or via reinforcement learning, and a neural net can be trained to work extremely well. It's about to what extent can an AI practice something. The environment has to be resettable. You can start a new attempt. Efficient, a lot of attempts can be made. And rewardable. There's some automated process to reward any specific attempt that was made. The more a task slash job is verifiable, the more amenable it is to automation in the new programming paradigm. If it is not verifiable, it has to fall out from neural net magic of generalization, fingers crossed, or via weaker means like imitation. This is what's driving the jagged frontier of progress in LLMs. Tasks that are verifiable progress rapidly, including possibly beyond the ability of top experts, I e math, code, amount of time spent watching videos, anything that looks like puzzles with correct answers. And while many others lag by comparison, creative strategic tasks that combine real world knowledge, state context, and common sense.