Latest / In the Loop AI / This Week in AI: The Big AI Short Failed

Transcript

This Week in AI: The Big AI Short Failed · In the Loop AI

0:00
Josh: If you were alive in 2008, chances are you remember the market crashing,
0:04
Josh: a catastrophic failure of the housing market that bled into the entire economy,
0:08
Josh: and it wiped out a lot of entities.
0:09
Josh: And if you'll remember, there was one person in particular who profited a tremendous
0:13
Josh: amount off of this happening, and his name is Michael Burry. He made $100 million.
0:17
Josh: Personally, collectively, he made a billion dollars off of shorting this market crash.
0:23
Josh: He saw what was right early, and he doubled down on it, and he made a killing.
0:26
Josh: In fact, so much so that it became a movie that you've probably seen named The Big Short.
0:31
Josh: Now, Michael Berry recently has published a new position in recent SEC filings that we uncovered.
0:36
Josh: And Ejaz has been all over this over the last couple of weeks,
0:39
Josh: tracking the positions, trying to understand why he's making these things.
0:42
Josh: And if this bet is correct in predicting the next big bubble to pop,
0:46
Josh: which is seeming to be AI.
0:48
Josh: So he predicted the housing bubble in 08. He made a billion dollars.
0:52
Josh: He's predicting another bubble in 2025. five, is he going to make another billion dollars, EJS?
0:58
Ejaaz: The short answer is, I don't think he will. I think he's going to get blown out.
1:03
Ejaaz: So Michael Burry is back. He is short around $300 million worth of Nvidia shares.
1:09
Ejaaz: So he's shorting, at this point, the richest, wealthiest company in the world.
1:13
Ejaaz: And that is a big sign, basically saying that I think the air bubble is popping.
1:17
Ejaaz: So the question is, is it?
1:19
Ejaaz: Let's go through his thesis, Josh. I'm going to simplify his post here.
1:24
Ejaaz: Which is basically GPUs run the AI world, right?
1:28
Ejaaz: Trillions of dollars have been spent by Frontier AI labs collectively to train state-of-the-art AI.
1:34
Ejaaz: But the thing with these GPUs, Josh, is that they have a lifespan,
1:39
Ejaaz: right? They don't last forever, right?
1:41
Ejaaz: And he estimates that the top Frontier labs.
1:46
Ejaaz: Overestimating or artificially boosting the lifespan of the GPUs that they purchase.
1:50
Ejaaz: And you might be like, well, that sounds boring. Why is that interesting?
1:53
Ejaaz: Well, the thing is, if you have assets on your balance sheet,
1:57
Ejaaz: that factors into your stock price.
1:59
Ejaaz: So the point that he's making is all these big companies are artificially boosting
2:03
Ejaaz: the lifespan of the GPUs that they've purchased so that it inflates their stock price.
2:08
Ejaaz: And therefore, the real stock price is actually a lot lower.
2:12
Ejaaz: In fact, he says at the bottom of this tweet, by 2028, Oracle will overstate
2:16
Ejaaz: earnings by 26.9%, Meta by 20%, etc.
2:20
Ejaaz: And so his big short, Josh, is on the behemoth that is supplying all these Frontier
2:25
Ejaaz: AI labs, NVIDIA, saying that when the bubble pops, and he thinks the bubble's
2:29
Ejaaz: going to pop now, he will make a heck of a lot of money on this trade.
2:34
Josh: So is he saying the bubble's going to pop now, or is it going to pop in 2028?
2:37
Josh: Because he's referencing those numbers in 2028.
2:39
Josh: So I'm kind of curious how he's framing this is this like because i'm now i'm
2:44
Josh: thinking about our positions um is this something that he predicts is going
2:47
Josh: to happen in 2028 or is this more of a short timescale position if
2:51
Ejaaz: We remember from his original short i think he held a position for about a year,
2:56
Ejaaz: Or at least like the better part of a year. So I think he probably estimates
3:00
Ejaaz: that it's going to happen sometime within that time frame.
3:03
Ejaaz: But I have more evidence as to like why he might be thinking that the bubble is bursting.
3:08
Ejaaz: Have you heard of what a neocloud is, Josh?
3:12
Josh: The name is familiar, but I don't understand quite what that means. So please fill me in.
3:17
Ejaaz: Okay, so think of like AWS or any of these other typical cloud providers.
3:21
Ejaaz: But they're specifically focused on providing GPUs and data center pipelines for you.
3:27
Ejaaz: So think of like an AWS just for AI compute. That makes sense?
3:31
Ejaaz: And so what he did was he looked at the major NeoCloud providers and he looked
3:38
Ejaaz: at their stock prices over the last week, Josh, and they were down.
3:43
Ejaaz: So he upped his position and was like, okay, this is the bubble bursting. I'm gonna lay my claim.
3:49
Ejaaz: There's an issue with this thesis, Josh. I think he's dead wrong.
3:53
Ejaaz: And I'm about to walk you through four reasons why he's wrong.
3:57
Josh: Okay. So then just to kind of like wrap my head around this,
3:59
Josh: the NeoClouds are kind of infrastructure for people who don't want to build
4:02
Josh: their own infrastructure.
4:03
Josh: Kind of like you described AWS. If you want to get a data center,
4:06
Josh: but you don't want to build a data center, you offload that responsibility to a NeoCloud.
4:10
Josh: And he thinks that the cost that these companies are marking down these NeoCloud
4:14
Josh: prices at, particularly the GPUs is much higher because the depreciation happens
4:19
Josh: on a shorter timescale than the companies are writing off.
4:22
Ejaaz: Correct. And NeoClouds is one part of the picture, but it's a great example
4:28
Ejaaz: to lead with because so many frontier AI labs like Microsoft OpenAI actually
4:33
Ejaaz: pay these NeoClouds billions of dollars.
4:36
Ejaaz: Like Microsoft just signed a $19 billion contract with a NeoCloud provider called
4:40
Ejaaz: Nebius, which kind of sent its stock price going up.
4:43
Ejaaz: We talked about it on a previous episode. so like it's a
4:46
Ejaaz: good example to kind of like lead with but the
4:49
Ejaaz: take here is that these neoclouds are
4:52
Ejaaz: basically as you said overestimating the life cycle of these gpus and therefore
4:57
Ejaaz: they are wrong except that's not the case at all and i present to you my first
5:02
Ejaaz: counter thesis which is he's wrong about the two to three year depreciation
5:06
Ejaaz: cycle it's actually more like five to six in some cases even eight years.
5:11
Ejaaz: These GPUs are used for more than just training AI models.
5:15
Ejaaz: They are used for inference, and they're sometimes used just for general queries
5:19
Ejaaz: and distribution for these AI models, right?
5:22
Ejaaz: But the most staggering kind of example comes from the top doc,
5:26
Ejaaz: the top neocloud called CoreWeave, who kind of shows the opposite to what Michael
5:31
Ejaaz: Burry is estimating, which is, there are people booking up GPUs, Josh.
5:37
Ejaaz: Two quarters in advance, six months in advance.
5:40
Ejaaz: And these aren't the latest GPUs. These are GPUs that are like five years old,
5:46
Ejaaz: three years old, in some cases, even longer.
5:49
Ejaaz: So if you dig into the details of all these different NeoClouds,
5:53
Ejaaz: you'll start realizing that it's the opposite of what Michael Burry is estimating.
5:57
Ejaaz: One, these GPUs have a very long lifecycle. And two, they're being used for
6:01
Ejaaz: way longer than people expect for really important things within AI.
6:06
Ejaaz: They are oversubscribed, they are overutilized.
6:09
Josh: This is interesting because you're getting like these really two counterpoints
6:12
Josh: here where Michael Burry's subjective take where he's like, no, this is wrong.
6:16
Josh: And then you have this objective number, which is CoreWeave.
6:18
Josh: And they're saying, wait a second, we're actually selling out these old hard
6:21
Josh: drives or these old GPUs.
6:24
Josh: Two quarters in advance. So how do we kind of piece together who's right and
6:29
Josh: who's wrong through this?
6:30
Josh: Because it seems like the market demand, and intuitively it makes sense too,
6:33
Josh: that there is no shortage of people who want to generate tokens.
6:36
Josh: And even if you are paying a little extra in premium in terms of cost per kilowatt
6:40
Josh: for those older GPUs, it's probably still worthwhile because the amount of money
6:44
Josh: you could build on top of that is so large.
6:46
Ejaaz: Forget about GPUs. Even the CPUs are being used for all the AI distribution
6:50
Ejaaz: and coordination. I mean, look at this, AMD CPU tab has just gone up in this
6:55
Ejaaz: week's projected earnings, right?
6:58
Ejaaz: But to answer your question, it's like, okay, well, who's right here, right?
7:02
Ejaaz: I want to kind of, before I address that point, I want to look at another actor
7:07
Ejaaz: within this whole circular economy that we like to talk about,
7:10
Ejaaz: right? Which is the Frontier AI Labs, right?
7:13
Ejaaz: Google, who is an established company, they just had their first $100 billion
7:17
Ejaaz: revenue quarter, are using eight-year-old GPUs to like fund their whole thing.
7:22
Ejaaz: So again, another data point saying that I don't think these GPUs are old and or not useful.
7:28
Ejaaz: And then if you look at NVIDIA themselves, they also have crazy amounts of demand.
7:33
Ejaaz: They're oversubbed for years in advance for their Blackwell GPUs and all their
7:37
Ejaaz: new GPUs going forwards.
7:39
Ejaaz: So then the last actor that I would want to look at is where does all the demand come from?
7:45
Ejaaz: Like who's buying these GPUs? Are they satiating demand?
7:48
Ejaaz: And if you look at every other company like Google, who has a ton of end users,
7:53
Ejaaz: OpenAI with chat GPT users, et cetera.
7:55
Ejaaz: Every insider take points to there's not enough GPUs to supply all of this demand
8:01
Ejaaz: that they're getting from the apps that they produce, from Sora,
8:04
Ejaaz: from a bunch of these different things.
8:05
Ejaaz: There's just too much demand. And in many cases, they don't have enough scale
8:11
Ejaaz: to even kind of like meet this demand.
8:13
Ejaaz: Jensen Huang had a conversation with TSMC, which is like their main
8:18
Ejaaz: chip manufacturer asking them to increase the rates of production by 50%.
8:22
Ejaaz: So the point I'm saying is Michael Burry, I think is dead wrong here.
8:26
Ejaaz: He's definitely underestimated demand and he's underestimated the life cycle
8:29
Ejaaz: of these GPUs. They are more valuable than he thought.
8:33
Ejaaz: And the old GPUs are worth more than they were worth when they sold originally.
8:38
Ejaaz: They're still selling within 5% of their contracts, which is just crazy.
8:41
Josh: So maybe I have to ask a pretty dumb question or seemingly
8:44
Josh: dumb at least and it's like why why is that so important to the bubble how companies
8:50
Josh: factor off losses in gpus over time is this some real big existential threat
8:55
Josh: to me it feels like a small part to a bigger picture and not quite as big as
9:00
Josh: something that would totally blow out an entire market that's been built so far the
9:04
Ejaaz: Biggest budget spend for any major ai company is on GPUs.
9:11
Ejaaz: That's where all the hundreds of billions are going for from Microsoft.
9:14
Ejaaz: That's where the majority of OpenAI's $1.4 trillion that they've committed over
9:19
Ejaaz: the next five years is going to.
9:20
Ejaaz: It's all to these compute providers. It's all to these neoclouds.
9:23
Ejaaz: It's all to creating their own chips. It's all to Jensen Huang's pocket.
9:28
Ejaaz: And so most of the CapEx bubble comes from GPUs.
9:32
Ejaaz: So if the bubble is going to burst anywhere, it's going to be from over-leveraging
9:35
Ejaaz: on GPUs. Do you want to know something crazy.
9:38
Josh: Josh? What's that?
9:38
Ejaaz: Even though hundreds of billions have been spent this year alone on GPUs,
9:45
Ejaaz: the companies who have spent it haven't even made a dent on their balance sheet.
9:50
Ejaaz: Remember, we're talking about Google here. We're talking about Microsoft here.
9:53
Ejaaz: They make hundreds of billions of dollars in net profit per year.
9:57
Ejaaz: They have plenty of flush cash and they're spending it on machinery that they think is going,
10:03
Ejaaz: you and I maybe don't see on the enterprise side or on the end consumer side,
10:07
Ejaaz: but they are obviously seeing it.
10:10
Josh: Okay. That makes sense. So the depreciation really is a big factor because it's
10:13
Josh: kind of artificially inflating numbers across the board.
10:16
Josh: It's from the people who are borrowing it, the people who are issuing the GPUs.
10:19
Josh: There's lots of high inflation that he's assuming is being baked into this.
10:22
Josh: And eventually that inflation kind of fizzles its way out through either some
10:25
Josh: really aggressive event or just over time degrading, which makes sense.
10:30
Josh: Okay. I think i think i'm up to speed on this i think i get his case
10:34
Josh: but what seems a little bit different this time is that
10:36
Josh: the first short that was made was in 2008 it
10:39
Josh: was against basically the entire stock market and particularly the housing market um
10:43
Josh: that was not growing nearly as fast as the ai market is that's like shorting
10:47
Josh: the internet and if you shorted the internet before 1999 or after like there's
10:53
Josh: this very narrow window to be correct without getting blown out because what
10:56
Josh: a lot of people don't realize is in order to short a company you have to borrow
10:59
Josh: money against from somewhere else.
11:01
Josh: And that borrowing comes with a premium. You have to pay a certain percentage
11:04
Josh: interest in order to loan the money.
11:06
Josh: If he's not right on this, or even if he's wrong by a couple of months,
11:10
Josh: with the rate that the market is accelerating, it feels like it's a very difficult
11:14
Josh: thing to get right because...
11:16
Josh: You not only have to time it right, but you can't get blown out by the appreciation
11:21
Josh: of things as fast as they're growing.
11:22
Josh: Like I've never seen a hockey stick of an industry steeper than this one.
11:26
Josh: So it seems like this is a really tough position to be bearish in or a tough
11:30
Josh: time to bearish in at least.
11:32
Ejaaz: Yeah, it's a it's incredibly risky trade to make.
11:36
Ejaaz: And I mean, the cherry on top of all of this, Josh, is that he paid the price pretty dearly.
11:43
Ejaaz: You sent a message to our group chat this morning and I had to read it like multiple times.
11:48
Ejaaz: Michael Burry, this morning of our recording, announced that he is closing down
11:54
Ejaaz: his fund, the very same fund which took out this $300 million short on NVIDIA.
11:59
Ejaaz: He was down on his position pretty massively. He overstated his means.
12:04
Ejaaz: And there's one line here, Josh, which kind of sums it all up.
12:07
Ejaaz: He goes, my estimation of value in securities is not now and has not been for
12:13
Ejaaz: some time in sync with the markets.
12:16
Josh: Tough. Okay, so this connects the dots for me because I saw the beginning and I saw the end.
12:20
Josh: I was like, okay, I know he goes out of business, but I do not know why or how
12:24
Josh: he gets there. And I think this probably connects the story.
12:26
Josh: Don't, again, to the point yesterday, don't bet against the optimist,
12:29
Josh: dude. He had absolutely cooked. So where does that leave us then?
12:32
Josh: So therefore, it is not a bubble and he was wrong or therefore it is a bubble
12:37
Josh: and he was wrong with timing.
12:38
Ejaaz: No, I like where you're sniffing josh because that something still seems a bit
12:44
Ejaaz: off right it's like well okay if it's not gpus what could it potentially be also so.
12:49
Josh: He he blew up his position but that doesn't ruin the hypothesis right like just
12:55
Josh: the prices moved against him but it doesn't prove the thesis wrong is that right
13:00
Ejaaz: It doesn't prove the thesis wrong on a long-term horizon because no one knows
13:05
Ejaaz: whether any of these crazy spends are gonna be crazy in hindsight,
13:10
Ejaaz: But for now, in the short term, all the fundamentals show that there is adequate
13:14
Ejaaz: demand for the GPUs and there are end users that are willing to use these AI
13:18
Ejaaz: products that require these GPUs.
13:20
Ejaaz: It actually tells us the opposite story, which is there are not enough GPUs
13:24
Ejaaz: in the world right now, old and new, that can satiate the demand for all the
13:28
Ejaaz: AI products that are being served today, right now at this moment.
13:31
Josh: Okay. Yeah.
13:33
Ejaaz: But one thing that could signal a constraint, Josh, is energy.
13:40
Ejaaz: The, put simply, the U.S. energy grid isn't up to par to supply energy to all
13:46
Ejaaz: these GPUs so that they can do the job.
13:48
Ejaaz: In fact, there's this clip that I have here from Satya Nadella of Microsoft
13:52
Ejaaz: where he goes on to basically say he has hundreds of millions of dollars worth
13:57
Ejaaz: of NVIDIA GPUs that are collecting dust in his fair weather data centers because
14:02
Ejaaz: they don't have the energy to supply this.
14:04
Josh: That seems uh troubling and i
14:07
Josh: think this is to a point that is being made
14:10
Josh: a lot where there are these gluts where
14:13
Josh: they are and how they show themselves is going to be the thing for debate and
14:18
Josh: this isn't the only hot clip we got this week there is this hysterical clip
14:22
Josh: um of satya nadella ceo of microsoft kind of talking about his relationship
14:27
Josh: with microsoft and i want to play this out before we give some commentary on
14:30
Josh: it because i think of all the clips i watched this week. This was one of my favorite.
14:33
Ejaaz: In our case, the good news here is OpenAI has a program which we have access to.
14:38
Ejaaz: And so therefore, to think that Microsoft is not going to have something that's- What level of.
14:43
Josh: Access do you have to that?
14:44
Ejaaz: All of them.
14:44
Josh: You just get the IP for all of that. So the only IP you don't have is a consumer
14:48
Josh: hardware. That's it. Oh, wow. Okay.
14:53
Ejaaz: Interesting. That's so good. That's crazy.
14:55
Josh: And this comes off the back of Satya saying in a previous interview
14:58
Josh: when asked about open ai he goes we are like above them we
15:01
Josh: are beside them we are around them they have this full total
15:04
Josh: controlling entity over open ai
15:07
Josh: and this is it's like the delivery is
15:10
Josh: hysterical the reaction is hysterical but it also it
15:13
Josh: shows a lot of interesting dynamics between these large companies because during
15:18
Josh: this interview satya also made mention of the fact that people who are building
15:23
Josh: their moats around individual models are very fragile and very brittle in the
15:27
Josh: sense that they are one company copying their model away from being worth significantly less.
15:34
Josh: And we kind of saw that earlier this week with our Kimmy K2 episode.
15:36
Josh: I highly recommend you watch it if you haven't, because
15:39
Josh: It shows how fragile being a frontier model can be if that's your only business.
15:46
Josh: And in the case of Satya and Microsoft, I mean, it seems like they have all
15:50
Josh: the leverage in the world.
15:51
Josh: If they want to fork the OpenAI code base right now and create their own chat
15:54
Josh: GPT, they can do that. They own all of it.
15:56
Ejaaz: Yeah, I've said this before, but I think the biggest winner of OpenAI is Satya
16:02
Ejaaz: Nadella. He has struck the best deal.
16:06
Ejaaz: OpenAI recently restructured their entire company. We actually did an episode on that.
16:10
Ejaaz: Feel free to check it out. But one of the major takeaways from that was,
16:15
Ejaaz: Satya owns 27% of OpenAI. And now he is able to engage in any model provider, not just OpenAI.
16:22
Ejaaz: He doesn't have exclusive rights anymore. He can kind of like flirt with other
16:26
Ejaaz: AI companies. And he did.
16:27
Ejaaz: He engaged with like Amazon on a bunch of different things.
16:31
Ejaaz: And he just owns all the IP until something like 2030 or 2032.
16:36
Ejaaz: So he is one of the smartest chess players.
16:39
Ejaaz: Honestly, unexpectedly for me, I didn't think I thought Microsoft was kind of
16:43
Ejaaz: like boomer in this sets, but he's navigated it beautifully.
16:47
Ejaaz: And one thing that I also need to give him kudos for is he has one of the biggest
16:51
Ejaaz: moats when it comes to enterprise.
16:53
Ejaaz: Like remember, Microsoft software is a really good consumer grade product,
16:57
Ejaaz: but is mainly kind of making their money from all the enterprise stuff,
17:00
Ejaaz: from Copala and stuff like that.
17:02
Ejaaz: So beautifully executed from Satya here.
17:05
Josh: I love this. Man, Satya, he's crushing it.
17:07
Josh: And it's funny because you don't think of Microsoft as a serious player in the world of AI at all.
17:12
Josh: I don't use any of their software. I don't use Copilot. In fact,
17:15
Josh: he was asked about Copilot because Copilot was the single AI coding agent.
17:19
Josh: It owned 100% of the market share.
17:21
Josh: And now it's taken down to 25%. And he said, this is a great thing because now
17:25
Josh: I still own 25% of a market that is now 10 times the size of what I previously
17:30
Josh: owned 100% of. So for him, he's playing the positive sum game.
17:33
Josh: He's very aware of the market dynamics at play. And for the cost of what was it, $10 billion?
17:38
Josh: He got the largest shareholder position
17:41
Josh: of open ai which was rumored to ipo at a trillion dollars
17:45
Josh: so they not only see the financial upside but they get all the intellectual
17:49
Josh: property most importantly the code base to either reverse engineer for their
17:53
Josh: own wants or to just clone and use for whatever they want it for it they have
17:58
Josh: it and it is a true chess move by satya bravo nicely done microsoft as
18:03
Ejaaz: Of today as we speak today at open ai's 500 billion dollar valuation uh microsoft,
18:10
Ejaaz: Microsoft stake is worth $136 billion.
18:15
Josh: Not a bad deal. But that's not the only big news. We have exciting news on the ChatGPT front.
18:22
Josh: This is what I'm most excited about because I use ChatGPT every day and this
18:26
Josh: is hopefully going to change the way I use it. Maybe we'll see when we talk about it.
18:31
Josh: The news this week is that ChatGPT and OpenAI, they launched GPT 5.1,
18:36
Josh: which is a whole new upgrade to GPT-5.
18:39
Josh: How different it is, we're not really sure.
18:42
Josh: I know, EJ, as I kind of read through the highlights, it's warmer in terms of
18:46
Josh: its like sentiment by default. So it'll be a little bit nicer,
18:50
Josh: which is bizarre because I thought we were going, we were trying to go the opposite
18:54
Josh: direction, like it's a little too nice.
18:56
Josh: And they added some safeguards that we'll get into in a second.
18:58
Josh: There's also much better instruction following, which is interesting because
19:02
Josh: oftentimes when you instruct GPT-5 to do things, it doesn't always do them exactly as you want.
19:07
Josh: So like if you say always respond in six words, it will oftentimes respond in
19:12
Josh: more or less than six words. It just doesn't quite understand the instructions.
19:15
Josh: And then one of the really fun things that I saw was they changed the different
19:20
Josh: types of personalities.
19:21
Josh: So you can choose between personalities on how you engage with the model.
19:25
Josh: And the previous ones were default, friendly, and efficient.
19:29
Josh: And now there are some newly added ones, which are professional,
19:33
Josh: candid, quirky, nerdy, and cynical.
19:36
Josh: So in the case you want to personalize ChatGPT to be more like those models,
19:41
Josh: well, now you have the option to do so with GPT 5.1.
19:45
Josh: So if I'm sick of it saying, you're so right, I agree, that's a great idea,
19:50
Josh: that's a great question, you could just say, cut that out, and it'll actually
19:52
Josh: listen for the first time, which is really exciting.
19:55
Josh: EJS, did you find anything interesting from GPT 5.1 after going through everything?
19:58
Ejaaz: Yeah, I spent the last like 12 hours playing around with it.
20:02
Ejaaz: Can I just, I'm going to play Bad Cop for a bit, Josh. Cool. I don't care.
20:06
Ejaaz: Like this is kind of like a nothing burger 0.1 update.
20:11
Ejaaz: And it's honestly like I've come to expect more from open air.
20:14
Ejaaz: So I'm kind of surprised that they've come through with this release.
20:17
Ejaaz: On the point of different personalities, I think that's super helpful.
20:21
Ejaaz: Because speaking for myself, I think that the current model is too agreeable. It's too sycophantic.
20:28
Ejaaz: And kind of Sam heard this feedback and the original version of GPT-5 was less
20:34
Ejaaz: sycophantic, but then a bunch of people were like, I want it to be more friendly.
20:37
Ejaaz: I want it to be more appeasing. And so he's kind of like flittering back and forth.
20:41
Ejaaz: This option gives a lot more flexibility for the end user, for me.
20:44
Ejaaz: I'm like, I want someone that's maybe a bit more candid or a little more cynical
20:48
Ejaaz: when I'm kind of doing my research and then maybe a little more friendly when
20:51
Ejaaz: I talk about more personal stuff, right?
20:52
Ejaaz: So the pliability is cool.
20:55
Ejaaz: What I will say is like, you kind of were able to do
20:58
Ejaaz: this before you could just go into settings user personalization
21:01
Ejaaz: and type in a prompt this i guess makes it easier
21:04
Ejaaz: because you just gotta click a button and maybe that helps a lot of people
21:07
Ejaaz: um the other cool thing i guess is like uh when it comes to efficiency it uses
21:12
Ejaaz: fewer tokens and gives you more output so you get a higher quality answer for
21:17
Ejaaz: less energy for less compute which is probably going to drive down costs of
21:20
Ejaaz: accessing this type of aid this type of Frontier AI, which is super cool.
21:25
Ejaaz: But aside from that, it's kind of like, why did they do this?
21:28
Ejaaz: One of the critiques I saw, Josh is like,
21:31
Ejaaz: it seems kind of rushed. Like there's no benchmarks, no API release.
21:36
Ejaaz: A bunch of the dev tools look a little sloppy. It's almost like they kind of rush this.
21:40
Ejaaz: It's not, there's no formal kind of like Sam's doing a live stream around this.
21:44
Ejaaz: It's nothing around like, you know, hey, here's how it compares against other
21:47
Ejaaz: models. They just kind of like rush this out. And I'm kind of confused.
21:51
Josh: Yeah. I suspect the public sentiment will start to reshape itself around how this goes.
21:55
Josh: Like when, when Apple releases iOS 26, it
21:58
Josh: was a huge release with liquid glass and all this cool new improvements and
22:01
Josh: then 26.1 is an incremental improvement where it adds a lot
22:04
Josh: of new features nothing really groundbreaking but the product gets light slightly
22:07
Josh: better i think we can kind of view these mid-tier updates
22:10
Josh: as that where they're just small incremental improvements like
22:13
Josh: now it's a little more efficient like you said there's this
22:15
Josh: new dynamic thinking time which is what they call like
22:18
Josh: smart thinking and it allows it to think faster
22:21
Josh: think more efficiently use less words that
22:25
Josh: are fillers so it needs less tokens to operate like the chart that we're
22:27
Josh: seeing and that just allows for one the
22:30
Josh: model to think more because it uses less tokens to think
22:33
Josh: and then two it just creates a little more efficiency unlock for open
22:37
Josh: ai's gpus to open them up to do other things so i assume this update is probably
22:41
Josh: for both people for the consumer and for open ai in terms of efficiency and
22:45
Josh: slightly better product is it this amazing novel breakthrough it doesn't appear
22:49
Josh: like it at all in fact it's just like a very marginal improvement but it's something it is like
22:54
Josh: It is something, and I suspect we'll probably get more updates like this,
22:57
Josh: where it's like, oh, well, it does this like one thing a little bit better.
23:00
Josh: Do you notice it? Maybe like one in every 50 prompts, but it's not a big deal.
23:04
Josh: And that's probably where we are for a little while until maybe Gemini 3.
23:07
Josh: I don't know. We have, there are murmurs of Gemini 3 coming down the pipeline from Google.
23:11
Ejaaz: I've been waiting for it for a month now.
23:14
Josh: Yeah, that could hopefully blow things out of the water. But we're at this weird
23:16
Josh: kind of stagnant period where we haven't seen those step function improvements
23:19
Josh: in AI models in a little while now.
23:21
Ejaaz: I mean, like the speculation that would be like we're kind of grinding to a
23:25
Ejaaz: halt and that exponential up, that S curve where we see AI improving massively
23:30
Ejaaz: and it's curing cancer may not be quite there yet.
23:34
Ejaaz: But hey, listen, in the meantime, there are several parties that are super bullish
23:38
Ejaaz: open AI. One of them is Masayoshi-san of SoftBank.
23:43
Ejaaz: In this news, he sold his entire stake of NVIDIA worth $5.83 billion to buy OpenAI stock.
23:52
Ejaaz: It's hilarious because all that money that is investing in OpenAI is only going
23:55
Ejaaz: to get spent on NVIDIA GPUs.
23:59
Ejaaz: Just hold NVIDIA. Like, why are you doing that?
24:02
Josh: Masa's really, he's been making some questionable decisions lately.
24:05
Josh: He sold NVIDIA too early and he missed the whole bubble he missed the whole
24:08
Josh: like he's just been making some strange decisions here he
24:12
Ejaaz: Wants to own 5% of NVIDIA do you know how much that would be worth today? I cannot believe,
24:19
Ejaaz: So like, you know, someone like kind of like tweeted this. They were like,
24:21
Ejaaz: you know, someone should look into what happened after he sold his entire NVIDIA stake in 2019.
24:26
Ejaaz: Just for reference here, the GPT 3.5 that went viral across the internet released
24:32
Ejaaz: two and a half years later.
24:33
Ejaaz: So pretty, pretty insane thing there. And then in final news,
24:38
Ejaaz: Josh, tell me about this. This looks super exciting.
24:40
Josh: The Google DeepMind team is back with some pretty awesome news.
24:44
Josh: We are very fond of the Google DeepMind team in particular here at Limitless.
24:47
Josh: They are very good at building real world physics in a digital manifestation,
24:52
Josh: which I think is a really important thing to be good at in a world where we're
24:56
Josh: training robots to be good at physical stuff, but they have not quite had the
25:01
Josh: training time to get good at physical stuff.
25:03
Josh: And what I love is that we've seen a lot of these projects coming out of the
25:06
Josh: DeepMind team. And today we got a
25:10
Josh: An agent that is capable of navigating these 3D virtual worlds.
25:13
Josh: So we saw previously there was a few models that can generate the worlds.
25:16
Josh: Well, now SEMA 2, which implies there was a SEMA 1 that I was blissfully unaware
25:19
Josh: of, allows these AI agents to actually navigate these virtual worlds.
25:24
Josh: So now not only can Google generate the virtual games, they literally look like
25:28
Josh: video games in this in this video, but they can also have the characters navigate
25:32
Josh: these complicated experiences, understand how things work, kind of piece things together.
25:37
Josh: It's amazing to see because it feels like as i'm watching
25:40
Josh: this video it is how a toddler would play
25:43
Josh: a video game it's kind of they're slowly moving it's figuring
25:46
Josh: it out you could see kind of reasoning in real time and navigating
25:49
Josh: these spaces and it's important to understand as you're watching this everything
25:52
Josh: that is on this is generated by ai because previously
25:56
Josh: these look like trailers to video games and those those video
25:59
Josh: games would have required thousands of people who are game developers lots
26:03
Josh: of time on the game engine and all of that is
26:05
Josh: done now fully by ai so seeing this announcement
26:08
Josh: was really exciting for me because i'm like one i just love video games and
26:11
Josh: two like oh my god wait these ais not only are they able to generate the environment
26:15
Josh: but now these ai models are able to actually engage with the environment and
26:19
Josh: what downstream effects does that have on training humanoid robots and the like
26:23
Josh: which we saw with the tesla episode last week yeah
26:26
Ejaaz: I mean yeah to your point um we see a real life implementation of this with
26:31
Ejaaz: tesla tesla Tesla has a world model which they use in their robots,
26:37
Ejaaz: which is the Tesla cars and automated self-driving and the future optimist robots.
26:42
Ejaaz: And what's cool about this is, Josh, one of the main things that makes your
26:46
Ejaaz: AI models super cool or really smart or intelligent is data.
26:51
Ejaaz: There's a scarcity of data. There's a scarcity of rich human data.
26:56
Ejaaz: Like they don't necessarily see what we see through our eyes or hear what we hear.
26:59
Ejaaz: These simulated environments basically synthetically create this data and enhance
27:04
Ejaaz: any of these AI models, right?
27:06
Ejaaz: And so Tesla kind of gets smarter by replaying a scenario where it almost crashed
27:11
Ejaaz: over and over and over again in simulated environments until it gets it right.
27:15
Ejaaz: And then that just gets transported to every single production ready Tesla that
27:20
Ejaaz: anyone is driving so that if they found themselves in a similar scenario, they avoid it.
27:24
Ejaaz: And that's just like one small use case for like what these robots will eventually
27:28
Ejaaz: be benefiting from. So such a super cool update.
27:32
Josh: Yeah. And if you scroll down actually to post EJs, you'll see a lot of similarities
27:35
Josh: to what we spoke about with Tesla, where you could see the model kind of reasoning in plain English.
27:39
Josh: You could see, oh, we equipped the pickaxe and copper mine. And then,
27:42
Josh: okay, like I'll equip the pickaxe and it walks you through its chain of thought
27:45
Josh: and then not only that but the post below it shows that it's self-improving
27:48
Josh: so the more time it spends cycling through the more time it spends playing games
27:52
Josh: the better it gets and these are early implementations of what what we kind
27:56
Josh: of imagined ai would eventually be which is this self-recursive loop where it
27:59
Josh: can improve without outside information and it's really exciting to see it happening
28:04
Josh: in the virtual world, because I guess that's the only place it can make sense
28:08
Josh: where it costs much less.
28:09
Josh: You can't harm anybody if a robot goes rogue, but it can learn without the inputs
28:13
Josh: of other people. So that is the exciting thing about this, I think.
28:17
Josh: And I'm really excited for the DeepMind team, man. I hope they keep going.
28:20
Ejaaz: Well, dude, like, do you remember our conversation with Logan Kilpatrick,
28:24
Ejaaz: the head of Google's AI studio?
28:26
Ejaaz: If you haven't seen that episode, definitely go and check it out.
28:29
Ejaaz: One of his main takeaways or one of my main takeaways is all of Google's AI
28:34
Ejaaz: byproducts feeds off of different models within the same suite.
28:38
Ejaaz: So if their video model learns something cool, they can transpose that onto
28:43
Ejaaz: their LLM, which is like in Word, so that it gets recursively smarter over time.
28:47
Ejaaz: I just think it's like fascinating and world models kind of like manifest all
28:52
Ejaaz: of these different things in one particular simulated environment.
28:56
Ejaaz: So, so cool. Kind of hard to wrap my head around sometimes, but because it looks
29:00
Ejaaz: like a video game to your point but but awesome.
29:01
Josh: Yeah it's you have to break this gap between like your normal understanding
29:06
Josh: when you see these videos and think like oh my god wait this is not normal this
29:09
Josh: is actually all ai generated um but with that we have concluded all the fun
29:12
Josh: new updates from this week it was a
29:14
Josh: i'd say a medium week it wasn't anything crazy there was no crazy outliers but
29:18
Josh: a solid week overall every week there's so much stuff going on and the progress
29:21
Josh: i feel like we're almost getting immune to it how fast things are moving um
29:25
Josh: but i hope everyone found this interesting if you did please don't remember
29:29
Josh: to share with your friends, like, subscribe.
29:31
Josh: Ejaz, I know you were manifesting some sort of send-off today. Do you want to share?
29:35
Ejaaz: Yes. Similar to Michael Burry, you shouldn't be shorting the biggest podcast bubble in the world.
29:41
Ejaaz: And Limitless is at the forefront over here. We're delivering you all the latest AI news.
29:46
Ejaaz: And just like the trade that Burry made that he lost $300 million in,
29:51
Ejaaz: you don't want to be shorting the best leading AI podcast in the world.
29:55
Ejaaz: Give us a five-star rating.
29:56
Ejaaz: Subscribe if you're not. What was the stat, Josh? 80% of listeners?
30:00
Ejaaz: 83%? How many of them don't subscribe? Over 80%.
30:04
Ejaaz: Over 80. That's nuts. It helps us out massively.
30:08
Ejaaz: We are currently in the top 30 in the technology podcast across Spotify,
30:12
Ejaaz: Apple, and wherever you listen. Help us get to the top 10. It would mean a lot. What do we say?
30:16
Josh: Don't ever bet against the optimists. And by subscribing and by being a part
30:20
Josh: of this, you are part of the optimists.
30:21
Josh: And we are going to move forward and things are going to progress.
30:25
Josh: And just don't bet against the optimists. So thank you for being optimistic
30:29
Josh: with us and being on this journey with us and we'll be back next week with a
30:33
Josh: whole new slew of episodes so thank you for watching and we'll see you then