B (1:12)

That's going to be a new meta.

A (1:13)

Yes, yes. You can't fake that. It's amazing. Also, it's just funny because, you know, it's effectively getting caught on a hot mic. But I was joking. I was like, of all the things that you could say on the hot mic before you sit down. Oh, okay. We're actually recording. His is just completely reaffirming everything we know about Ilyas Hutsquare. It's just completely the same. Like, okay, he is a true believer. It's not like he was sitting down and being like, newark cash, we gotta go on my private plane. I just sold so much secondary. It's crazy what's going on with this stuff. Like, if people really think this AI thing's gonna pan out, I'm making billions of dollars. I'm cashing out. I don't believe any of this stuff is real. No, he wasn't caught on a hot mic like that. His hot mic moment is like, wow. It's exactly like science fiction, everything.

B (2:01)

It's all real.

A (2:02)

It's all real.

C (2:02)

Yeah.

A (2:03)

Which is just iconic. Tyler, did you have any other takeaways from your Speedrun? You're listening to it at 5x, right? Does he pop the scaling bubble? Does he give a bearish take about Is it over at any point?

C (2:15)

So I wouldn't say he's like anti scaling, but he does kind of give this interesting take, which he basically says that AI companies, like, there's too few ideas for the amount of companies and for the scale that we're at. You can think of AI progress as being in these kind of distinct ages. Right. So he says 2012-2020 was like the age of research, where you're trying all these different ideas and the scale of things is very small. Right. To train the original Alexnet was like two GPUs. To do the original Transformer was like eight, maybe 64, but like, you know, very small amount of GPUs. Once we kind of figured out that transformers work, we entered this age of scaling and that's basically from 2020 to 2025. And now we're basically at this point where like, yes, you can keep scaling and models will get better, but even if you scale 100x, like, are we really going to get super intelligence? It'll get better on the benchmarks and they'll become more useful, but it's not like this. He doesn't think that just raw scaling alone is basically what's going to bring us there. I mean, this has been echoed by a lot of people. We still need a couple different kind of paradigms for this to work. The reason that Opus 4.5 was better is not just because they scaled pre training, it's scaling. Generally the scaling has gone from pre training and now it's rl.

A (3:28)

Yeah.

C (3:28)

And so we basically. We need to find another paradigm. And the way you do that is just doing, like, research. And so he talks about SSI as basically being this like, return to research.

A (3:37)

Return to research.

B (3:38)

Yeah.

C (3:38)

It's small kind of training runs, even though, you know, they only raised 3 billion, which is like, small compared to other research institutions. The fact that they're basically putting it all on these kind of. I mean, I don't know if they're moonshots, but they're these small training runs where they're doing experiments and then they're going to scale that up eventually. But they're not just basically trying to win the AI race by just scaling up and doing the same thing as everyone else.

A (4:01)

Yeah, yeah. They're trying to find a new. A way to actually bend the scaling curve, find a new scaling law, or find a new technology that like, they can scale against. I was thinking about Ilya's talk at Neurips last year. He pulls up this chart of the relationship between the mammals mass and the brain volume and it's a pretty linear graph. And so like the elephant is a lot bigger than the mouse and so it has a proportionally larger brain to its body volume. And it's this perfect, it's this perfect linear curve. I should just try and figure it out if I can maybe text it in. Boom. So basically the mammals have this very clear linear trend, but then the non human primates are a little bit higher up on the chart and they're just doing a little bit better. But then hominids, the actual humans have a different, there's a very distinctly different curve. It was making me think like maybe that's what we're supposed to see when we think about. Yeah, this, this when we say like straight lines on log graphs, when we say we are seeing scaling happen with the current architectures, which line are we scaling against? Are we actually scaling on the, on the human curve or are we waiting for divergence from that curve?

C (5:28)

Current scaling law scaling has taken all the air out of the room, right? Where like basically we have more than enough compute to try these like different ideas, but they're just all going straight into training the next big model using the next paradigm. And maybe it's slightly different, right? You have a different way of doing RL or whatever, but it is still fundamentally the same thing. Right. And he talks about maybe continual learning is really the better approach. Right. We've been in this era of like having a pre training thing for so long that we think of like AI is like you train this thing and then you release it and it's like done. And RL is like a little bit different now because there's this idea of post training and you can kind of integrate different.

A (6:02)

I also thought the interesting thing was with, with pre training you use the whole Internet so you don't have to decide anything. You're just applying this algorithm to just all the data, all the compute and there's no decisions. But then with rl, you have to decide, okay, we're putting in these math equations and we're maybe not putting in something else because we're actually creating the data and we're, it's not just this.

C (6:22)

Like, this is maybe why we see these kind of like models that are super well, they do super well in evals, but not so much.

A (6:29)

Yeah, some of the overfitting.

C (6:30)

And the reason is because the data that we choose is not the correct data because researchers are basically being reward hacked maybe into just solving for benchmarks.

A (6:40)

It's interesting to hear this, the conclusion is we need another breakthrough. And then simultaneously consensus be. But we're definitely going to get that breakthrough in the next decade.

B (6:53)

I feel like it echoes a lot of even what Mike New has been saying. We need new ideas. Yeah, totally saying this for, for months and.

A (7:00)

But it's way, it's way harder to predict the rate at which breakthroughs will arrive as opposed to like you can actually chart out, okay, the formation of capital, the time it takes to build a data center, how long it takes to, you know, manufacture a bunch of GPUs, rack them, run the training run. Like, that's much more predictable than like human came up with new algorithm that's sort of random.

C (7:20)

And he brings this up as, as the reason why you see companies doing this because it's just if you're raising money, it's so much easier to justify the raise by saying, we're going to buy this data center. Totally do this training run. It's going to cost exactly as much as this way. Yeah. And then model will be this good and then we can use it to monetize this way.

A (7:36)

Totally.

C (7:36)

Where if you're just saying like, oh yeah, we're just going to pay a bunch of like really smart researchers to do a bunch of research and then they'll figure something out. You can't really.

B (7:43)

Yeah. In some ways it feels like SSI is set up for like somewhat of a mini AI winter or like at least riding the hype cycle down. Because it doesn't sound like he's sitting there being like, we raised 3 billion and we're spending it in the next 12 months. It's like 2.9.

A (7:58)

Was that.

B (7:59)

No, not, not, not.

A (8:01)

No, no, no, no. That's the point. It's not. It's like equity. It's just sitting there. It's like he can clearly pull back.

B (8:06)

I'm going to give each researcher different teams, like a shots on gold.

A (8:11)

No, I love it.

B (8:11)

We're going to keep taking those shots until obviously he'd be able to raise like another $10 billion whenever he wants, especially if he has like a key breakthrough insight and they can be first to scale that. We're delighted by Google's success. They've made great advances in AI and we continue to supply to Google. Nvidia is a generation ahead of the industry. It's the only platform that runs every AI model and does it everywhere computing is done. Nvidia offers greater performance, versatility, versatility, fungibility than Asics, which are designed for specific AI frameworks. Or functions.

A (8:46)

That is a crazy thing to post.

B (8:47)

Crazy, crazy, crazy thing to post.

A (8:50)

Sometimes you get stuff.

B (8:51)

I don't know, boys. But having the largest company in the world sending tweets to the to defend their main product is not very reassuring.

A (8:57)

I feel like this would be so much better delivered. I actually don't have that much of a problem with the actual text here. This should be delivered by Jensen with some nuance in a conversational setting. It just hits a lot different when this is in at exactly 9am like clearly scheduled, clearly typed out in a document. It feels like a press release, which is just an odd, odd thing when, when it should be. This should be an answer to a question. Someone Bobby Cosmic in the chat was saying like, oh, the mainstream media is just now picking up on the Gemini 3 story and there's articles in the Wall Street Journal and other places saying like, oh, maybe Google's back. Like, you know, by Google. Like it's very exciting. And so Nvidia feels, feels the need to respond to that. But it's a lot different when it's actually a response instead of just like a we're putting out a press release like who knows why? As opposed to like Jensen saying like, well, since you asked, talk show host or news anchor or whoever, he's top podcast host, whoever he's talking to, Dwarkash, whoever he's talking to, maybe us, we'd love to have him. I can ask him that question. He can defend this here.

B (10:05)

Well, the timing seems important because they are coming under a huge amount of pressure right now. There was an article in Barron's this morning by Tae Kim. The headline is not what Nvidia's comms teams would have liked it to be. Nvidia says it's not Enron in private memo refuting accounting questions.

A (10:25)

That's a crazy thing to say.

B (10:26)

Let me get, let me get into the coverage. So Tay says a series of prominent stock sales and allegations of accounting irregularities have put Nvidia in the middle of a debate about the value of artificial intelligence and its related stocks. Now Nvidia is pushing back. In a private seven page memo sent by Nvidia's investor relations team to Wall street analysts over the weekend, the chip maker directly addressed a dozen claim made by skeptical investors. Nvidia's memo, which includes fonts in the company's trademark green color, begins by addressing a social media post from Michael Burry last week which criticized the company for stock based comp dilution and stock buybacks. Burry's bet against subprime mortgages before the 2008 financial crisis was depicted in the movie the Big Short. Of course, Nvidia repurchased 91 billion shares since 2018, not 112 billion. Mr. Burry appears to have incorrectly included RSUs, RSU taxes. Employee equity grants should not be conflated with the performance of the repurchase program, Nvidia said in the memo. Employees benefiting from a rising share price does not indicate the original equity grants were excessive at the time of issuance. That makes sense. Barron's reviewed the memo, which initially appeared in social media posts over the weekend and confirmed its authenticity. Burry told Barron's he disagrees with Nvidia's response and stands by his analysis. He said he would discuss the topic of the company, stock based comp and more details. Burry is of course now over on substack. He's charging $380 a year and if you are a perma bear, I can't. This is like Christmas coming early. Nvidia didn't respond to Barron's for request for comment, but they also responded to claims that the current situation is analogous to historical accounting frauds Enron, WorldCom and Lucent that featured vendor financing and SPVs. Unlike Enron, Nvidia does not use special purpose entities to hide debt or inflate revenue. Nvidia also addressed allegation that its customers, large technology companies, aren't properly accounting for the economic value of Nvidia hardware. Some of the companies use We've talked about this use a six year depreciation schedule for GPUs. Burry said he believes the useful lives of the chips are shorter than six years, meaning Nvidia's customers are inflating profits by spreading out deep depreciation costs over a long period. The TPUs equal bad for Nvidia take is up there with the dumbest, maybe worse than deep seek as it completely misses what actually happened in the last six weeks and I will remember who is who in the zoo. My view one demand for AI is bananas. No one can meet demand. Everyone is spending more. Google said just yesterday they have to double capacity every six months to keep up. Two scaling laws are intact. He's referencing Gemini 3. The flywheel is about to speed up. Somehow the mid curve crew thinks this is zero sum competition. None of this suggests that. If you think the race is hot now, wait until you see what comes out of large coherent Blackwell clusters. All the magic from the quote God machines is pretty much still hopper based. Lastly, a quick GPU TPU less than the cost and performance specs on the box aren't what you get in real life and Google is going to get a bat margin 2 double duh. What matters is system level effective tokens to watt to dollars and TCO. Nvidia GPUs have higher FMU because they're already embedded in workflows slash the ecosystem is massive. By the way, this is a good test if you have an opinion on this topic, but you have to look up FMU then perhaps curate better source.

A (13:46)

View what mfu mfu what I said fmu.

B (13:51)

The above effective token watt gap also likely widens with Ruben. Add in that Jensen can actually deliver volume in a tight market plus future flexibility multi cloud capable, programmable for paradigm shifts and he'll sell every GPU he makes for years. Google will too since everyone wants a second supplier and TPU is a fantastic chip. But this is as far from either or as it gets. The one benefit of this confusion is that it is likely to give Google a brief stint as the world heavyweight champion, the most valuable company I would guess the Midwest Put the strap on them in less than two weeks Put.

A (14:25)

A strap on them? What does that mean? Just like pile in. It seems like he's predicting that people will overplay the Nvidia bear take and overplay the Google opportunity Opportunity. And that will result in Google becoming the most valuable company in the world. And he uses the phrase put the strap on them in multiple in less than two weeks.

B (14:49)

According to today's Wall Street Journal, AI related investment accounts for half of GDP growth. A reversal would risk recession. We can't afford to go backwards. The article is how the US Economy became hooked on AI spending President Donald J. Trump unveils the Genesis Mission to accelerate AI for scientific discovery Today, Trump signed an executive order launching the Genesis Mission, a new national effort to use artificial intelligence to transform how scientific research is conducted and accelerate the speed of scientific discovery. The Genesis Mission charges the Secretary of Energy with leveraging our national laboratories to unite America's brightest minds, most powerful computers and vast scientific data into one cooperative system for research. The order directs the Department of Energy to create a closed loop AI experimentation platform that integrates our nation's world class supercomputers and unique data sets to generate scientific foundation models and power robotic laboratories. The Order instructs the Assistant to the President for Science and Technology to coordinate the national initiative and integrate an integration of data and infrastructure from across the federal government. There's one more note here on strengthening America's AI dominance. Trump continues to prioritize America's global dominance in AI to usher in a new golden age of human flourishing, economic competitiveness and national security.

A (16:05)

Yeah, I'm very interested to hear how, like, how the public private partnership actually works here. There was a time when every, basically every cool technology was coming out of darpa, coming out of the US Government. The US Government landed on the moon. And since then, I think a lot of people in technology have lost faith in the US Government overseeing the development of technology, even academia. I mean, people think AGI will emerge from a private C Corp. That's where people believe that the best work will be done. Give Ilyas Sutzkever. Give the best scientist $3 billion. Let him go cook. That's the Currently, this feels like somewhat of a rejection of that in some ways. There's obviously lots of different places where having AI resources, having science and technology resources within the government make a ton of sense. But it'll be interesting to see, like, where are the interfacing points between the two. Between the two categories? By default, I think most people in our audience in technology would say, hey, let's leave the, let's leave the space travel and the AI research to the private sector. Should we run through The Astral Codex 10 piece on trait based embryo selection? This is from Scott Alexander in Astral Codex 10. He says suddenly, trait based embryo selection when a couple uses so in 2021, genomic prediction announced the first polygenically selected baby. When a couple uses IVF, they may get as many as 10 embryos. If they want one child, which one did they implant? In the early days, doctors would just eyeball them and choose whichever looked the healthiest. Later, they started testing for some of the most severe and easiest to detect genetic orders, like disorders like down syndrome and cystic fibrosis. The final step was polygenic selection, genotyping each embryo and implanting the one with the best genes overall. Best in what sense? Genomic prediction claimed the ability to forecast health outcomes from diabetes to schizophrenia. For example, although the average person has a 30% chance of getting type 2 diabetes, if you genetically test five embryos and select the one with the lowest predicted risk, they'll only have a 20% chance. So you get a 10% bump there. That's nice. Since you're taking the healthiest of many embryos, you should expect a child conceived via this method to be significantly healthier than one born naturally. Polygenic selection straddles the line between disease prevention and human enhancement. In 2023, Orchid Health, founded by Noor, who we've had on the show. Enter the field. Unlike Genomic Prediction, which tested only the most important genetic variants, Orcid offers whole genome sequencing, which can detect the de novo mutations involved in autism, developmental disorders and certain other genetic diseases. Critics accused GP and Orcid of offering designer babies, but this is only true in the weakest sense. Customers couldn't design a baby for anything other than slightly lower risk of genetic disease. You're basically just selecting out of what you already got. They're not editing the genes, they're merely sequencing them and then allowing you to select. These companies refused to offer selection on Traits, the industry term for the really controversial stuff like height, IQ or eye color. Still, these were trivial extensions of their technology. And everyone knew it was just a matter of time before someone took the plunge. Last month, a startup called Nucleus took the plunge. They had previously offered 23andMe style genetic tests for adults. Now they announced a partnership with Genomic Prediction for focusing on embryos. Although GP would continue to only test for health outcomes. You could forward the raw data from GP to Nucleus and Nucleus would predict extra traits including height, bmi, eye color, hair color, adhd, iq, and even handedness.

B (19:52)

And it's worth noting that Nucleus is now being sued by Genomic Prediction, even.

A (19:58)

Though they have this partnership.

B (19:59)

I'm assuming the partnership is no longer. We can ask, but I'm assuming it's no longer because one of the GP's co founders left, left the company predictions to join Nucleus.

A (20:11)

Interesting.

B (20:12)

And allegedly turned off all the security cameras.

A (20:18)

Is that metaphor, or is that actually.

B (20:20)

The lawsuit alleges that he turned off all the security cameras on his.

A (20:24)

That's not a metaphor for like, you know, sharing a Google Drive of PDFs.

B (20:29)

It's his last day at work.

C (20:30)

Okay.

B (20:30)

And he was allegedly like rounding up.

A (20:35)

Okay. So he turns off the cameras. Allegedly. And the implication is that maybe he was rummaging around, like literally taking documents or something like that. That's at least what the timeline is.

B (20:44)

What the lawsuit alleged.

A (20:46)

Okay. Wow.

B (20:47)

People at Nucleus were emailing the former co founder at his old email address, evidence of them violating the. The agreement that they had. So anyways, it's very, very, very, very messy. We can ask.

A (21:03)

Yeah, there's like four or five companies involved in this and all of them.

B (21:07)

Are controversial because this is the most, I think the most controversial, probably, like, category that you can be in.

A (21:13)

Yeah, it's certainly up there. And also there's just like the. There's just. It's so Easy to throw. I mean, in the same way that people are throwing Enron at, at Nvidia. Like it's so easy to throw Theranos at any biotech company that's not, you know, that's accused of anything. And also biotech, it's like it's, it's pretty hard to understand the underlying science. It's not, it's not a, it's not as popular as, okay, like, does the website work? Does the business make money? You know, what's the cash flow like? It's way more complicated and so it does attract even more attention. So one of the other companies in the space is herasite and astral Codex 10 continues here. They enter the space with the most impressive disease risk scores yet, an IQ predictor worth six to nine extra points, and a series of challenges to competitors whom they call out for insufficient scientific rigor. Their most scathing attack is on Nucleus itself, accusing its predictions of being misleading and unreliable. Let's start with the science and then move on to the companies to see if we can litigate their dispute. In theory, all of this should work. Polygenic embryo screening is a natural extension of two well validated technologies, genetic testing of embryos and polygenic prediction of traits in adults. So genetic screening of embryos has been done for decades, usually to detect chromosomal abnormalities like down syndrome or simple gene editing disorders like cystic fibrosis. It's challenging. We've talked about this before. You need to take a very small number of cells, often only five to 10, from a tiny protoplacenta that may not have many cells to spare, and extract a readable amount of genetic material from this limited sample. But there are known solutions that mostly work. And so the companies that we're talking about today aren't necessarily doing like the fundamental lab equipment development, building the machine, figuring out how to sequence data from the first. It's more about the analysis that happens.

B (23:12)

On top of the results and the recommendations. And the recommendations, which is probably, which I would say is the most controversial part of this.

A (23:21)

I don't know that any of them are recommending, hey, we think you should pick this baby. They're more just saying, like, we think that according to the data, this baby might be.

B (23:30)

But if you're giving somebody risk fact, if you're giving, if you're.

A (23:33)

Yeah, but that's not a recommendation. If I tell you this car is 700 horsepower and does 0 to 60 in two seconds, and this one does 800 horsepower and does 0 to 60 in 2.4 seconds. This one's faster in a straight line. This one's faster on the curves. And then, like, you pick, like. I didn't make a recommendation. I just told you the stats. If a company engages in malpractice, for example, plagiarism, providing products they should know are bad to customers, etc. Is it water under the bridge if they can clean up? That's obviously a reaction to my question was, you know, is there a redemption arc in his mind? Somebody says, volkswagen can answer this question really well. I think that's because Dieselgate. I just feel like the next. The next turn of discussion needs to be, okay, we tested the models, we tested the data, we tested the claims at a higher level of rigor, I guess.

C (24:22)

Sejuan Mala is also accusing Kian of using a Chad filter.

A (24:25)

This happened before. So when Keon came on the show maybe six months ago, Growing Daniel accused him of using a Chad filter and went super viral. And I was kind of like, oh, that's. That's. I don't know. I don't know how to even respond to that. That's a very silly claim. I have no idea if this is real. I can't tell at this point on a zoom call at this resolution. What do you think? Do you think this is real? Are you guys just cracking up? Because everyone. Does everyone think it's real?

C (25:04)

I don't think that he used a filter.

B (25:06)

I don't think he used a filter.

A (25:07)

I don't think he used a filter either. He just grew a beard, and I think he's just been mewing maybe.

C (25:12)

Maybe he's just photogenic.

A (25:13)

Yeah, it is possible that he just. He just, you know, flexed his jaw muscles and, like, you know, has low body fat. I don't know. I feel like it would be extremely high risk to run a. A chin augmentation filter down for a second.

C (25:27)

I mean, because you.

A (25:29)

You know, that's what happens, right when you're using, like, the Snapchat filter or like the TikTok filters. Like, sometimes they pop in and out, and if they pop out, like, Yerk done.

B (25:37)

Gotta get the Nucleus test for the Gigachad test and the publicize the results.

A (25:45)

Everyone's cracking up in the studio having a wild time anyway.

B (25:48)

This is actually insane, apparently, according to X. I don't know if this is true, but the robbery that took place yesterday in which an armed thief posed as a delivery driver and robbed somebody for $11 million of Ethereum and Bitcoin, was Locky Groom that was targeted.

A (26:09)

Whoa. What an armed thief posing as delivery guy finessed his way into the $4.4 million Mission District home shared by investor Lockheed Groom. Yes, Sam Altman's ex boyfriend and another tech investor named Joshua.

B (26:22)

Oh, okay, so it was not Lockheed, but Joshua.

A (26:29)

Gary Tan posted the footage, panicked enough to delete it minutes later. Crypto security experts are now saying what everyone thinks. Self custody is great until someone shows up your door with a fake UPS label and a Glock. San Francisco's tech leader about to hard pivot into vault custody. Private security zero public flexing. Because this heist wasn't random. It was a warning shot.

B (26:47)

Very chatgpt written Mario Nofal. But anyways, very sad.

A (26:53)

We will be back on Friday. Friday for Black Friday. We have a fantastic lineup of a bunch of different entrepreneurs. E commerce founders, brand builders, some of.

B (27:05)

The most savage operators. E commerce operators.

A (27:07)

It'll be a lot of fun in the world.

B (27:09)

Cannot wait.

A (27:09)

It's gonna be a great time.

B (27:10)

A lot of friends. Have a wonderful Thanksgiving. We are thankful for each and every one of you. Thank you for being a part of this and we'll see you Friday.

A (27:18)

Goodbye. Cheers.