- Emergent Behavior
- Posts
- EB-1: SemiAnalysis - Transcript - Part 3
EB-1: SemiAnalysis - Transcript - Part 3
Ate A Pi
April 13, 2024
Transcript: SemiAnalysis - Part 3
💼 Dylan Patel: there will be misaligned AI. So I need to research now or I need to know what I need to do or else someone else will take this market over or what have you, right? So I think like, sure, in a year from now, you're gonna be able to buy a way better GPU, right? For, but like, who cares, right? Like, in March of last year, if you wanted to rent an H100, maybe you could get it in June and June, you'd get it for a way higher price than you get it today. But like, who cares, right? Cause I couldn't have done any of the work I did in the meantime, right?
👩🎤 Ate-A-Pi: Wow.
👩🎤 Ate-A-Pi: Wow.
👩🎤 Ate-A-Pi: Wow.
👩🎤 Ate-A-Pi: Wow.
💼 Dylan Patel: So it's like, you know, sure, it's cheaper, sure. Compute. I mean, like this is the age of like, yeah, Moore's law kind of was stronger historically, right? Like you would get two X decreases in manufacturing costs and then you'd have architecture costs on top of that or performance increases on top of that. Right. And it was like, and people didn't delay purchases, right? Like they just purchased when they needed it. And every time, you know, the, the cost. Havd right. And every time the performance more than doubled, right. You, you got, you got more demand, right? Because let's go, wow, all these new use cases got on lock.
👩🎤 Ate-A-Pi: Wow.
💼 Dylan Patel: Right. So today people are like generative video was like so fucking expensive to inference. It's like, yeah, well, wait till people improve the algorithms by 10 X and the chips improved by 10 X. And now I can have like customized video content that is perfect quality stream to my eyeballs for like a penny a minute. Right. Like, you know, what, you know, it's like, it's this is going to happen. And it's like, you either like build it or you get eaten. Right. And that's, I think it's a race. Right.
👩🎤 Ate-A-Pi: Wow.
👩🎤 Ate-A-Pi: My gosh, that's insane. I cannot imagine a financial model where you're like, you have to budget kind of like 50%, 75% depreciation in your capital, in your built-up capital. My gosh, that's crazy.
💼 Dylan Patel: I don't know if it will be 5075, but certainly within four years, your GPU is worth a tiny fraction of what it used to be because of the new stuff that's out. I think that's tough, but that's not the only world where that's the case. There's plenty of other industries where consumables were a thing, you use it and you lose it, kind of things. In general, the history of semiconductors were this way.
👩🎤 Ate-A-Pi: Wow.
👩🎤 Ate-A-Pi: Right.
Right, right, right.
Right, right, right.
👩🎤 Ate-A-Pi: All right.
💼 Dylan Patel: especially like early 2010s, buying a new phone every year made sense or every two years, every three years, right? Maybe nowadays, like people don't really care about like one year versus three years, but like, you know, you look at the Apple headset, right? Like you buy the Apple headset today and you're like, holy crap, guess what, a year and a half from now, that headset's gonna be garbage compared to what Apple releases, right? And what Meta releases. And it's like the same applies to the Quest 3 and whatever Meta drops next, right? And it's gonna be like that for AR slash.
👩🎤 Ate-A-Pi: Right. Oh, yeah. Yeah.
👩🎤 Ate-A-Pi: Right.
💼 Dylan Patel: VR slash mixed reality, whatever you want to call it. And it's going to be like that for AI, right? And the thing is, the people who buy Apple headset potentially are going to be more productive, are going to have more entertainment value sooner, right? And same applies to the GPU, right? They're going to be more productive and they're going to be able to launch a business or learn more things or put out research and that's going to advance humanity, right? If everyone waits, then no one's going to build anything. And so, thankfully the financialization of AI has not happened yet.
👩🎤 Ate-A-Pi: Yeah.
👩🎤 Ate-A-Pi: Yeah.
💼 Dylan Patel: And like even Google's like, fuck it, build it and they'll come. Right. So, you know, like, so like, there's no one that's really like, Oh, no, what about the ROI? It's like, no, we'll build it and they'll we'll make money. Right. And that's the attitude that I feel like, you know, opening, I has, I feel like that's the attitude that deep mind has that's the attitude that meta has this attitude that all these open source people have, they're like, they're like, I don't care about the ROI. I'm just going to drop the weights and then people will teach me what I'm doing wrong, right? Like, you know, that's, that's like, it's the beautiful thing, right? Not, not just the weights, but the data pipelines or what have you.
👩🎤 Ate-A-Pi: All right. All right, all right, all right.
👩🎤 Ate-A-Pi: Alright. Right.
👩🎤 Ate-A-Pi: All right.
👩🎤 Ate-A-Pi: Right, right, right. Yeah. Right, right. That's amazing. All right. I'm going to ask, I'm going to divert again. So I had like a basically a post on ASML this week where I was like, oh, you know, they
💼 Dylan Patel: I wanted to call you an idiot for that, sorry.
👩🎤 Ate-A-Pi: no, you should go ahead and call me an idiot. So why was I wrong? So I had, basically, in summary, I had a post where I was like, my engineering intuition is like they can be beaten, that you have 4,000 suppliers, they have 20 million lines of code. It doesn't look, it looks like the Boeing Lockheed kind of engineering milieu of the last century before SpaceX came in, and the intuition is that they can be beaten.
So why was I wrong?
So why was I wrong?
💼 Dylan Patel: Um, so not that, not that it can't be beaten because other people, there's people working on it, right? Like, like China's trying to make an, uh, UV tool with a particle accelerator for the light source. Right? Like that's super cool. But like, you know, if you, if you, you know, I think, I think like one of the things you said in there was like, not everything needs to be cutting edge or like, I don't think people quite understand how freaking complicated this tool is, but it's also the only way we can get it. Right. So, so you have to have a light, right? You have to have light.
👩🎤 Ate-A-Pi: Yeah. Right.
👩🎤 Ate-A-Pi: Yeah. Right. Yeah.
👩🎤 Ate-A-Pi: Yeah.
💼 Dylan Patel: that we can bounce off with mirrors and focus, right? Because that's the only way we can pattern a photo resist, right? Maybe there's a different way to manufacture, but no one's gotten close on cost with any other way, right? People have tried nano-imprint lithography, which you can think of like as stamps, right? Way less accurate and has way horrendous yields, right? Manipulating light is simply the best way so far.
👩🎤 Ate-A-Pi: Yeah.
👩🎤 Ate-A-Pi: Alright.
👩🎤 Ate-A-Pi: Right.
👩🎤 Ate-A-Pi: All right.
💼 Dylan Patel: And a lot of the cool stuff that people are trying to work on is also just manipulating light, but it's manipulating light and hitting a chemical that chemically changes when light hits it, right? That's photoresist. And you do it in a certain pattern, right? And that's basically, you know, and then you do that, you know, alongside a bunch of like putting metal down, putting oxides down, putting whatever down and etching, right? Now you iterate hundreds of steps, thousands of process steps actually, to manufacture a chip, because a chip is transistors, contacts, and then dozens of metal layers.
👩🎤 Ate-A-Pi: Yeah.
All right.
All right.
👩🎤 Ate-A-Pi: Alright.
💼 Dylan Patel: That's on-off switches perfectly arranged to make sand think, like in simplistic terms. How do you put something down at a 20 nanometer pitch? One, visible light is hundreds of nanometer wavelength. Ultraviolet light is still hundreds of nanometer wavelength. So you step down to, oh, OK. And then there's this weird gulf where you can't do a lot of stuff.
👩🎤 Ate-A-Pi: Right.
👩🎤 Ate-A-Pi: Yeah.
👩🎤 Ate-A-Pi: All right.
💼 Dylan Patel: It's hard to really like, you can't really like bounce it off mirrors. So for years and years, we only had DUV lithography, 193i. We figured out you could put it through water and it would kind of focus. I'm kind of simplifying, but basically you'd focus it more, right? It's NA and I don't want to explain all that stuff. Anyways, there's Asianometry on YouTube if you want to learn how a lithography tool works in an entertaining manner. Anyways, the EUV happens and it's like, EUV gets researched, right? We have DUV for...
👩🎤 Ate-A-Pi: Yeah, he...
💼 Dylan Patel: literally two decades before EUV comes out. And the whole time EUV is being worked on. And it's like, there's this massive chasm that we physically cannot do anything in. There's 193 immersion lithography. Before that, it was 193 not immersion lithography. So there wasn't really a change in wavelength or anything. It was just the immersion capabilities, which was still like hard, but like, anyways, there's a massive gulf for 20 years where we're just doing more and more and more immersion lithography, stacking that step so we can get finer and finer pitch. And then along comes EUV. But...
👩🎤 Ate-A-Pi: All right.
👩🎤 Ate-A-Pi: Right.
👩🎤 Ate-A-Pi: All right.
💼 Dylan Patel: to make EUV, the only way we know as humanity to make EUV light, right, at 13.5 nanometers, which is a Goldilocks zone for wavelength of light, is by dropping a tin particle, blasting it with a laser, so it spreads out and then blasting it with a laser again, so then all these excited photons shoot out, collecting them all on a mirror, and then focusing it, sending it through a bunch of lenses and mirrors to focus it more and more, putting it through a mirror that has a pattern on it, but mask, right?
👩🎤 Ate-A-Pi: Yeah.
👩🎤 Ate-A-Pi: Right.
💼 Dylan Patel: and then bouncing it off a more and then hitting the wafer finally, right? That's the only way we know how to, right? Now, is there a simpler way? Potentially, I look for anyone to even theoretically come up with a way to do this better and more accurate because the next step is just high NA EUV, right? Which has all these cost concerns, like as you mentioned, right? Boeingification, right? Well, high NA EUV is the next step after EUV and it's incredibly difficult, right? And it's like, people also don't understand, like there's, I wish I could share my screen and show you like a video of some of these things, but like...
👩🎤 Ate-A-Pi: Alright, alright.
👩🎤 Ate-A-Pi: Right.
👩🎤 Ate-A-Pi: Alright.
💼 Dylan Patel: Yeah, I talked about the laser, that's the source, right? That's called a source. That's engineered in San Diego by a company called CYMR that ISML owns, vertically integrated. But that tool, that itself, right, is droplets of tin at the perfect frequency, two lasers synchronized perfectly, this crazy expensive collector of mirrors to collect this light source at a humongously high power. Oh, so there's a laser in it that hits the tin, right? That's a CO2 laser, I think, if I recall correctly, from Trump, right? Like it's like...
👩🎤 Ate-A-Pi: All right.
👩🎤 Ate-A-Pi: All right.
💼 Dylan Patel: There's a tremendous amount of complexity, but how else do we create a 13.5 nanometer wavelength light or something smaller? Right. Cause that's what's better. And today we have no clue. The future is like particle accelerators and stuff like that. And I guarantee you that's even more complicated than the tin source. Right. And then you have the, the optic system, right. Which is all these lenses and that lens, the lenses are made by Carl Zeiss. And that's a tremendously complex supply chain in and of itself. But.
👩🎤 Ate-A-Pi: All right.
👩🎤 Ate-A-Pi: All right.
💼 Dylan Patel: Is there a more sophisticated way to make lenses and focus light and stuff? Probably maybe. I don't know of one. Maybe someone can figure it out, but that's done in a joint venture with ASML. And then there's the reticle system, which is basically that mask. Where, cause when you print a chip, you don't just expose the whole wafer and move on. You do sections of the wafer at a time because you can only ever expose a certain amount of area at once and you only generate enough photons from your source. There's tremendously complex.
Source has a hard time making enough photons to make it to the way for some, so they have to, you know, you only do a portion of the way for that time, right? And it's, you know, through a reticle. And the wafer, right, is moving at nine Gs in one direction and the reticle is moving at nine Gs in the other direction and they are constantly in sync in sub-two nanometer precision, right? And so like, yeah, that's a fucking complicated system with thousands of parts. And there's other parts on this system too, but it's like, how else do I expose
Source has a hard time making enough photons to make it to the way for some, so they have to, you know, you only do a portion of the way for that time, right? And it's, you know, through a reticle. And the wafer, right, is moving at nine Gs in one direction and the reticle is moving at nine Gs in the other direction and they are constantly in sync in sub-two nanometer precision, right? And so like, yeah, that's a fucking complicated system with thousands of parts. And there's other parts on this system too, but it's like, how else do I expose
👩🎤 Ate-A-Pi: Yeah. All right.
👩🎤 Ate-A-Pi: Right, right.
👩🎤 Ate-A-Pi: Indeed.
💼 Dylan Patel: hundreds of wafers an hour at that precision. And I have no clue what exists in the physical realm that we can pattern chips at that pitch, at that speed and at that cost, right? So people are like, holy shit, that's expensive, right? It's a $200 million machine. The next one, it's $180 million machine. The next one is $400 million, right? The high NA, the next generation of EUV. Yeah, that's fucking expensive, but the $20 billion fab has many of those tools.
👩🎤 Ate-A-Pi: Indeed, indeed.
👩🎤 Ate-A-Pi: Right.
👩🎤 Ate-A-Pi: All right.
👩🎤 Ate-A-Pi: Right, right, right.
💼 Dylan Patel: And by the way, the lithography tool is only like 30 to 35% of the cost of the FAP, right? And then there's a bunch of other stuff in there too, etch deposition, you know, all sorts of different steps, epitaxy, you know, all sorts of stuff that is also involved in making a chip. And yet I'm holding a phone with 20 billion transistors in just the logic processor and then 16 gigabytes of RAM, which is another billions of transistors, right? And each of those cost millions of a penny to manufacture.
👩🎤 Ate-A-Pi: Yep.
👩🎤 Ate-A-Pi: All right.
💼 Dylan Patel: So is it really inefficient? Is it really a disgusting supply chain? Or is it like crazy complex and somehow, the industry has halved costs and now it doesn't have cost anymore, but it's still reducing cost every year, right? And people say Moore's law is dead, like, yeah, it is not shrinking as fast as it used to, but transistors are still getting cheaper. That's also another lie that people sometimes spread as transistors aren't getting cheaper. They are, they're just people who are saying they aren't are using stupid math that isn't true.
👩🎤 Ate-A-Pi: All right.
👩🎤 Ate-A-Pi: Mm-hmm.
👩🎤 Ate-A-Pi: All right.
💼 Dylan Patel: Right? Like you can just look at the wafer costs and how many transistors, blah, blah. But like transistors are getting cheaper still. And, and I know of no other industry in the entirety of humanity that has reduced costs as fast as semiconductor manufacturing. Right? Um, there is no other industry that has, that has been at this clip, uh, that has gone from one to a hundred, a hundred billion for the same price in just five decades. Right? Um, you know, car manufacturing garbage.
👩🎤 Ate-A-Pi: True.
👩🎤 Ate-A-Pi: Right, right, right.
💼 Dylan Patel: Horribly inefficient. Tesla's so inefficient. SpaceX, yeah garbage. Okay, wow, they have the cost of launching a rocket to space. They take it to one 10th. Yeah, semiconductor industry does that in their sleep. They did that in five years, six years, right? You know, so like, no, like this is the most efficient thing that people have ever done. That's the coolest thing people have ever done. So when people slander, you know, my beloved semiconductor industry, and not to say like ASML is not like inefficient in ways, like there's definitely bureaucracy and inefficiency, like
👩🎤 Ate-A-Pi: Oh my gosh. Right. Alright, right.
💼 Dylan Patel: I'm sure it could be done better, but if it could, people would do it because the semiconductor industry is so valuable and there's so much money to be made by breaking the mold. But it turns out breakthrough innovation and even keeping up, like Nikon in Japan, Smee in China, there's an American lithography company, Canon, all these companies were in lithography and they got out competed.
👩🎤 Ate-A-Pi: Alright, alright, alright.
💼 Dylan Patel: by the beast that is ASML, right? Like it's not like, and why? Because ASML constantly reducing costs, right? So yeah, it's expensive. Yeah, it's complicated, but I think it's efficient as hell.
👩🎤 Ate-A-Pi: Right.
👩🎤 Ate-A-Pi: Mm-hmm.
👩🎤 Ate-A-Pi: Interesting, interesting. I don't know, I was wondering whether you would show up for the podcast. I was like, and cause you know, I don't take it personally, right? I'm like, part of the reason I put stuff out on Twitter is to get feedback on like ideas, right? Because you wanna have your ideas tested and people are absolutely brutal on Twitter, right? Like, which is just kind of fun, right? So.
💼 Dylan Patel: Sorry, had to rant. You really pissed me off with that tweet.
👩🎤 Ate-A-Pi: Um, and I, you know, so I'm like, ah, you know.
💼 Dylan Patel: Yeah, I almost quote retweeted you in real and was like, I'm gonna call you a fucking idiot. And then I was like, maybe I shouldn't. But I also know you do this, right? I remember the first time I messaged you about LK99. I was like, dude, this is clearly wrong. He's like, yeah, I'm just doing a bit. And I'm like, oh. And then I kind of got like, oh, he's doing a bit. Okay, okay. She ate a pie, he's doing a bit, yeah.
👩🎤 Ate-A-Pi: Yeah, I know, it's, it's...
👩🎤 Ate-A-Pi: Yeah, yeah. Yeah, so it's kind of like, cause I kind of wanna see what people say and I wanna get like, because oftentimes like when you say something which is not exactly right, you get a lot of feedback. And you like, Asianometry showed up man. Asianometry showed up on the thread and like, this is V1, we will get to the launch pad and then we will go into space and we will come back.
you know, Agent Autry. I was like, okay, dude, that's awesome, right? So I don't think, you know, I spent a bit of time following Jim Keller, right? And Jim Keller's viewpoint is basically that the system as it is right now is optimized for what it needs to do.
but you can change what it needs to do and then you can optimize it in a different way. So for example, he's like, if we had resilient logic functions, then you could allow a larger number of defects because currently I think he said like the allowable number of defects is like one per square centimeter. And if the defect numbers were higher, then you could have new design techniques on the chip. But because you have like this kind of like very small defect number that you're allowed,
and that's because your logic functions are not resilient, then you end up with this. So it's like, it's a cascade of issues, which maybe you can find like a way out of that cascade, that maze in, you know, if you had an innovation somewhere else in the stack, for example, in the design of the logic function, to make it more resilient, then you could allow more defects, not less, but more defects, which would then allow you to have like a different set of designs, which changes the parameters on which
the entire chip is being made, right? So, but the thing is, like, you know...
you know, Agent Autry. I was like, okay, dude, that's awesome, right? So I don't think, you know, I spent a bit of time following Jim Keller, right? And Jim Keller's viewpoint is basically that the system as it is right now is optimized for what it needs to do.
but you can change what it needs to do and then you can optimize it in a different way. So for example, he's like, if we had resilient logic functions, then you could allow a larger number of defects because currently I think he said like the allowable number of defects is like one per square centimeter. And if the defect numbers were higher, then you could have new design techniques on the chip. But because you have like this kind of like very small defect number that you're allowed,
and that's because your logic functions are not resilient, then you end up with this. So it's like, it's a cascade of issues, which maybe you can find like a way out of that cascade, that maze in, you know, if you had an innovation somewhere else in the stack, for example, in the design of the logic function, to make it more resilient, then you could allow more defects, not less, but more defects, which would then allow you to have like a different set of designs, which changes the parameters on which
the entire chip is being made, right? So, but the thing is, like, you know...
💼 Dylan Patel: Yeah, so I think there's a misnomer in what is the defect rate of semiconductor. Because a lot of people just talk about like, you know, there's this or that, right? So for reference, right? Five nanometer, right? Five nanometer chips, which have been in mass production since 2020, right? Extremely high yields. That's what H100 is made with. That's what all the iPhone since 2020 have had.
It's practically all the leading edge chips you can think of use 5 nanometer, whether it's PC chips from AMD or GPUs from Nvidia or Apple smartphone chips. You go down the list, TPUs, right? They're all 5 nanometer TSMC, right? For 5 nanometer, you can fit 140 million transistors roughly per millimeter squared. Right? Or in other words, right? Like...
It's practically all the leading edge chips you can think of use 5 nanometer, whether it's PC chips from AMD or GPUs from Nvidia or Apple smartphone chips. You go down the list, TPUs, right? They're all 5 nanometer TSMC, right? For 5 nanometer, you can fit 140 million transistors roughly per millimeter squared. Right? Or in other words, right? Like...
👩🎤 Ate-A-Pi: Mm-hmm.
💼 Dylan Patel: You can effectively have, you know, 24, I'm gonna do it in Apple units because that's easier. You can have, you can have, you can have 12.5 trillion transistors, right? And the defect rate on those transistors is infinitesimal. Right, it is actually irrelevantly small, right? And it just turns out when I make a $20 billion, a 20 billion transistor chip, right? Which is only 10, 100 square millimeters, 10 by 10 millimeters, right?
👩🎤 Ate-A-Pi: Uh-huh.
👩🎤 Ate-A-Pi: Right.
👩🎤 Ate-A-Pi: Wow.
💼 Dylan Patel: I can fit 624 of those on a wafer. When I make that chip, most of those work perfectly. And even the ones that don't work perfectly only have one or two transistors broken. People kind of disregard like... And the yields are in the 80% plus range anyways, for just functioning chips. When you look at functioning transistors, they're in the 99.99999999%. The TSMC 5-nanometer process...
👩🎤 Ate-A-Pi: Right.
👩🎤 Ate-A-Pi: Yeah.
👩🎤 Ate-A-Pi: All right.
👩🎤 Ate-A-Pi: Mm-hmm.
💼 Dylan Patel: has thousands of process steps, right? And it's gonna sound extreme, like why do we need thousands of process steps? Well, you need to make every layer of the chip and to make every layer of the chip, you have to put down material, you have to do lithography, you have to remove material, you have to clean it many times in the middle, you have to perfectly sand the wafer, it's called CMPs, chemical metal mechanical planarization. There's so many steps that need to be done. Maybe there's a way to do it simpler, right? But like, you know, even like novel stuff like Sam's Aluf's atomic semi, they're not targeting like these leading edge five nanometer chips, they're targeting, you know,
👩🎤 Ate-A-Pi: Right.
💼 Dylan Patel: decade or two back, but significantly faster and cheaper cost prototyping, it seems like. You look at the defect rates on leading edge manufacturing and they're infinitesimally small. You have a thousand process steps. If you were saying each one of them is at Six Sigma, people in manufacturing like Six Sigma, blah, blah. If Six Sigma was the manufacturing tolens for semiconductor manufacturing, no chips would work. No Apple iPhone chip would ever work.
👩🎤 Ate-A-Pi: I see. I see, I see, I see.
👩🎤 Ate-A-Pi: Right.
👩🎤 Ate-A-Pi: Right, right, right.
💼 Dylan Patel: coming out of the fab because there's that many process steps. There's that much precision. There's that many things that you have to make properly. Right. And I just talked about transistors. Then you talk about, you know, 14 layers of metal, connecting them all together perfectly. So it's like, you know, the, I think, I think there's a bit of like, people don't understand how like voodoo magic this shit is.
👩🎤 Ate-A-Pi: Right.
👩🎤 Ate-A-Pi: All right.
👩🎤 Ate-A-Pi: Right, absolutely. I'm okay, so I'm gonna, I'm gonna. Right.
💼 Dylan Patel: Like another example is GPD4, right? 1.8 trillion parameters. So if I took a piece of paper and I printed 50 on them, 50 parameters, right? Every single parameter in 16-bit, right? Which is like eight digits long or whatever, right? If you were to print every single one of those and then you printed every parameter in GPD4, you could literally line up the paper to the moon and back 23 times. I did the math on that because I was like, oh, I wonder. It's like, oh, and it's...
👩🎤 Ate-A-Pi: Right.
💼 Dylan Patel: And it has to multiply by each of those numbers twice, right? When it's doing, you know, in simplistic terms, when you're generating a single token for GPT-4, right? So it's like, holy shit, this shit is like very efficient or models are really inefficient or like, wow, this shit is just magic, right? Like you can go either way, but like that's how efficient like GPUs are, right? You know, people just don't like grasp, what is a teraflop? Oh shit, that's a lot of compute, right? Like, you know, oh, it's switching at gigahertz speeds, right? And there's trillions of millions of transistors.
👩🎤 Ate-A-Pi: Indeed.
👩🎤 Ate-A-Pi: Yeah, it is, it is. Right.
💼 Dylan Patel: You know, it's like, people lose sight of how fantastic what we've made is when they kind of get in this like, you know, oh, it could be simpler. Oh, it's not good. Blah, blah, blah.
👩🎤 Ate-A-Pi: It's amazing. I'm gonna do a lightning round to finish up. What is artificial general intelligence to you? What does it mean when someone says AGI?
💼 Dylan Patel: It's what's advantageous to use in the conversation. That's what they're using.
👩🎤 Ate-A-Pi: right on.
💼 Dylan Patel: It's better than humans in every way, I don't know.
👩🎤 Ate-A-Pi: better than humans in every way. Okay, if you had to put a date on like when AGI would be achieved, when do you think AGI would be achieved?
💼 Dylan Patel: Actually, that definition sucks. I think AGI in like 2028, but ASI, right? Artificial super intelligence, maybe like 2032 or something like that, right? Like it's gonna take a little bit longer. Which is, ASI is where it's better than humans in every way. Right? In terms of thinking.
👩🎤 Ate-A-Pi: Yeah.
👩🎤 Ate-A-Pi: Okay, so AGI meaning better than humans in every way. All right, and AGI being on par with humans in every way, something like that. Would you say that? Just generally in touch. Yeah.
💼 Dylan Patel: just generally intelligent, right? Like, you know, good and good and good, better than humans in some ways, right? Like better at recognizing cats and photos for sure, already happened, right? So it's AGI in that sense, right? It's generally intelligent in that sense. It's better at, you know, making poems or finding shit on the internet, but then it's way worse at other things, right? So I think there will be a continuum for AGI and people can have whatever definition they want there.
👩🎤 Ate-A-Pi: Alright, alright, alright.
👩🎤 Ate-A-Pi: Amazing, amazing. That's pretty much it for the show. I wanna thank you and I'm gonna end the show now and then I'm gonna switch off and I'm gonna go to the normal stream. So just give me a second. All right.
Reply