AI’s Economic Impact, Warren Buffett Discovers YouTube, Mistral Launches “Le Chonk” | Diet TBPN

I was leafing through GQ this morning. Tom Cruise is on the COVID It's a good edition. There's a funny anecdote in here that I wanted to share. I gotta find it. But they got the whole Tom Cruise spread, lots of photos. But they did a little back and forth interview with him, and he's a course bro. He dropped a course, a private course, and we gotta get access to this. The question from the GQ interviewer: You have mentored a lot of younger actors. Glenn Powell, has talked about watching a 6-hour instructional video about screen acting that you made, right? And Tom Cruise says, it was very basic. It was just something to have people understand, just to have a different concept of lenses and what the lens is and what the frame is, just the same way as an actor on stage.

Being a global superstar for dummies.

Yeah, basically. He made a 6-hour— that's a course. He made a 6-hour video. Just the same way as an actor on stage or in the development of cinema, where it started and what the— That frame means and how it—

Great success.

And how to understand that frame.

Thank you, team.

There's just different layers of storytelling, of structure, but also to understand certain things so that you can forget about it. What else do you tell younger people, other people in your business? I listen first. What do they want? It's not me. I'm more interested in what are their goals? What do they want? And it's not just younger people. This is me when I was a kid. I'm going, what are you interested in? They say, really? Yes. What is this movie about? What is this? There's a bunch of other fun things in here, but I thought it was funny that there's a secret 6-hour training video on how to be an actor from Tom Cruise. We got to get access to it. I feel like there's a lot of journalists that they get documents, they scoop things, they release things, but I've never seen anyone just be like, I'm putting down 10 racks for whoever can bring me the 6-hour Tom Cruise video.

Yeah.

I'm putting it behind a paywall. I'm gonna monetize it, but there's a bounty out. Like Kevin Roose has the Dario Amadei big blob of compute thesis from 2017, like this leaked PDF sort of on the scaling law and the bitter lesson basically. And I've never seen a journalist just be like, I know this document's out there. Like Eric Newcomer should just be like, bring me Founders Fund returns. I'll give you 5 grand if you got them. 'Cause it probably maths out at some point.

Maybe $5 million.

Maybe more, maybe more, yeah.

To risk your allocation in the next—

Yeah, I suppose that that would violate journalistic ethics to just put a bounty out for a document that you know exists.

I will bribe you for the truth.

And it probably really, like I was thinking about the various leaks that go out, like some of them violate NDAs, but if they're secretive enough and they're squirreled away, Via signal, like they probably don't trace back. They probably don't really trigger anything. Also, some of them just don't really matter. They you know for a company might not be that frustrated about a particular leak if it makes the company look good.

But well, speaking speaking of movies, the team saw movie yesterday. Ben, could we get a review? Oh, you didn't go.

Ben didn't. Wow, didn't come.

Yeah. Wow, Tyler, review.

I saw it.

Who saw it?

So we saw it.

Nick said that he was starting to look at his phone after like forty minutes.

Wait, what?

Nick did not enjoy it. I thought it was pretty good. So basically, the reviews are really bad, right?

Yeah.

And famously— well, not famously, but the box office is really bad.

Not all the reviews are bad. I saw Mark Andreessen gave it a positive review.

He did, yeah. And that was really one of the only very positive reviews I saw. But yeah, I think the budget was like $160 million. First weekend, it was like $8 million.

Cracking to like $8 million, yeah.

Pretty brutal. I thought it was good. It was very cool, like the framing of the movie. It's hard to reveal too much.

I thought it was very cool.

Did you go to film school? This critique is so nuanced. It's remarkable.

I thought it was—

I don't want to spoil anything. I don't want to spoil anything.

What was cool about it?

Just give us like 30 seconds of spoilers. I'll bleep it out.

Okay. So basically—

The bleep organ. The game decompilation thing is going strong. Ben Thompson was writing about it in Sir Teckery today. There were some more knock-on effects. He took it in a very interesting like legal direction about the implications of taking a video game, decompiling it, doing whatever you want with it. Obviously, IP licensors don't like that, but there's not a lot that you can— that you can do to control that. And obviously, open source models are coming. And there's a bunch of interesting knock-on effects. I saw a really funny one. They're getting funnier, these mashups. I saw one where someone was playing NHL hockey in Resident Evil. Did you see this one?

No.

It was really— it was really such a good mashup because the guy's just like on ice skates. And like flicking pucks at zombies. It was, it was pretty good. I was, I was into it. Again, probably not really something I want to play for 5 hours, but will I watch a 30-second video of it? Absolutely. But yeah, Polygon has been breaking this down. If you've been on X over the last few days, you probably noticed that things have gotten very weird when it comes to modded video games. There are clips of Call of Duty operators doing kickflips and other tricks from Skate, people playing Black Ops 1 zombie mode and Skyrim, Miles Morales swinging through Gotham. Here's another add to the list. Someone has gotten the Xbox version of Halo: Combat Evolved running in a web browser with multiplayer and split-screen support. And the other funny thing is that it's not— they didn't— like, apparently the Halo community doesn't like the PC release of Halo, so they ported the Xbox version to the cloud browser. It's like, you can really do anything you want now. And so Ben Thompson goes on to describe a little bit of what's going on. Games are written in human-readable languages, C, C++. Obviously there's scripting languages on top of that, but then they're compiled into machine code that's executable by the processor that they're intended for. For Halo, that's x86. You've always been able to disassemble machine code into assembly because assembly maps almost perfectly to machine code. The difficult part, which takes humans years, is decompiling assembly to C or some human-readable language that then you can actually modify. But LLMs are really good at that. And so is it legal? Is it legal? The story conflates 2 different AI capabilities. Fallout project appears to be a new game. There's no Fallout New York, although using the name Fallout is likely a trademark violation. I think we were joking about the lawyers being really excited seeing all these going out because it means full employment. Decompiling is not in and of itself illegal. Sega versus Accolade held that disassembling code is fair use if it is the only way to reach unprotected functional elements and there was a legitimate purpose. And Sony v. Connectix extended the same reasoning to reverse engineering the PlayStation's BIOS. However, that is likely where protections end. Code that is derived from decompiling is by definition very similar, if not identical, to the original code, which is covered by copyright. So even if you go and decompile a game and then you ship something that's like a mod, you're going to get a DMCA notice. And this is already happening. Like, something will go up, it'll get— it'll go viral, and then they'll get hit with a A takedown notice. This already happened. 2 hours later. Exactly. Is that me or is that—

Getting hit with a copyright.

2 hours later, you'll get hit by— yeah, copyright. So you have to take it down. There's a couple interesting things. So the actual value of these, like, decompiled games mashups, pretty limited because you basically just get a viral video and you can't really monetize it because you get a takedown notice. But the training rollout for the AI labs, like the training data that comes from having this fleet of all these different game modders go and spend a bunch of tokens that they're basically paying for, maybe it's a little subsidized, but they're basically paying for all this decompilation. And then also the labs probably have RL environments for this type of work. And then there's tools for this as well. Means that I would expect the next version of models that is trained on all the inference that's happening right now, because everyone right now— in the game modding community is like, let me go point GPT-Astra and Opus 5.5 at whatever game binaries I can get. So it's like, oh, I like Skyrim, I like Minecraft, let's go decompile those. But then they're also doing it with all the long tail. So all of that's going into the next set of training data. So you would imagine that the models get much better at video game development, which is cool.

Yeah.

He also just talks about like how things will change over time, similar to what's happened in video. AI seems to exacerbate a lot of trends including the shift to service-based games. So he talks about like in a world where if you send someone a disk or you send someone an executable binary, like they're just going to decompile it and be able to do whatever they want with it, maybe you want to stream that game. So you— this was— Google had an effort for this. I forget what that was called. But Google had the ability to basically stream you a 1080p—

Google Stadia.

Stadia, that's right. So they would stream you the video output and you would stream them the controller input and the game would run on the server. So you never saw any code, you just saw video coming back.

This is also how the Meta Quest Xbox Edition worked.

Oh yeah.

You were streaming the game in.

That's right, that's right. And I think Xbox has stayed with the cloud gaming a little bit longer, but gamers have always sort of pushed back on that. I don't like the latency. What if my Wi-Fi goes out? I'm just trying to play a single-player game, blah, blah, blah. And so it's possible that this is something where the game industry comes together and they're like, We really got to fight these hackers. We got to hide our code more behind more of a firewall, like that cloud streaming firewall. Or it could be something that they just like embrace and go the IP protection route. Yeah, it's hard to think that this doesn't go the way of generated text where like people still care about the posts and where they come from, effectively intellectual property, the instantiation of the actual product is less relevant than the story around it and the, and the people behind it and the, and the selling point of like, we're playing the same game. Of course people will do their individual stuff as well.

But anyway, what else we have going on, John? Well, I am so glad to be past that subject.

Because you just don't like video games?

Uh, I had enough. I had enough.

You had enough?

I had enough. What about racing simulator?

That's a video game.

That's the final frontier.

Yeah, you're gonna be able to race anywhere you want.

Mustafa says, coming in with a major white pill, AI won't take your job anytime soon. In 10 years, only 5% of what humans do will be replaced by AI. Let's see what he has to say about that.

They're doing new media over at MAI, the Humanist Review of AI. They made a magazine and they have a number of articles and they tapped MIT Economist and Nobel laureate, Nobel Prize winner in economics, Daron Acemoglu, to write a piece about how AI will diffuse. And it's interesting because in this one it's total white pill, but then he was on 60 Minutes and he's a little bit more black pilled. So we'll have to go back and forth. I think it's the same.

Wait, that's the same guy?

Yeah. We'll have to reconcile this. Okay. So let's see what Daron Acemoglu says in the Humanist Review. It's a longer piece, but Mustafa Suleyman summarized it here. He said, he argues that we need to stop building AI to replace people and start building it to make them better at their jobs. 52% of Americans are worried about AI's impact on their jobs. The fear is overblown and it's steering how we build AI. I like the 48% of just confident people. Like, good luck.

Good luck. I'd like to see you try. The type of work I do, yeah, I'm not in the training data.

I'm not— there's no RL environment for what I do. AI isn't in the productivity statistics yet. This was the famous line about the internet. It shows up everywhere except the productivity statistics. Most firms using AI aren't seeing real gains yet. Expect roughly 1.5% added to GDP over 10 years. Not a revolution. Electricity took decades to spread. There's so many of these—

The problem is like, so the problem is like going back to our conversation with Sean Frank yesterday from Ridge. This is a company that's over 10 years old right now, growing 50% this year, using AI across the entire business from demand planning to creative to a number of other things. And the business is performing pretty much as well as it ever has. The thing is, AI is impacting the whole business in different ways, in big ways and small ways, but you can't give AI credit for the growth because there's so many factors contributing, right? New products, push into retail, all these different things. And I think you can still say that maybe most firms aren't using it to see real gains. But when you say that electricity took decades to spread, I just don't view AI as it's not the kind of technology that takes a long time to spread when it works.

I agree.

And that's because we can use the internet, which pretty much everybody that's productive has access to, to spread AI.

Yeah.

So already just a few bullet points in and—

But the Nobel Prize-winning economist has another example to push back on. He says, if you don't like the electricity example, here's another one. Speech recognition. He said, turning speech into text has been an aspiration since the 1930s. And this is a harder one to grapple with than electricity because a lot of what was in place for AI is also in place— a lot of the things that you're saying about why diffusion should be faster for AI was true for speech recognition. So people have been trying to do speech recognition. You talk, the computer types. Since the 1930s, going on 100 years of pursuing this. It became a huge step closer to reality thanks to the work by American— by the American husband and wife team, James and Janet Baker, starting in the 1980s when they founded the iconic company Dragon Systems. Dragon truly revolutionized voice recognition, and it would not be an exaggeration to say that everything else that has followed in voice recognition since then is thanks to the Bakers. Their breakthroughs— their breakthrough was to eschew previous attempts to build speech recognition systems based on syntax and meaning and instead deploy machine learning techniques in order to build a pure prediction system which would map speech onto the best match text. Already in 1997, the year Dragon launched Dragon NaturallySpeaking, the first program that was capable of voice recognition from natural continuous speech The momentousness of this was obvious. So in previous software, including the previous product from Dragon, Dragon Dictate, you had to pause in between each word to let it process the gap between words. You no longer needed to do that with Dragon NaturallySpeaking. This was 1997. I can vouch for all of this because Darren Ismoglu, the writer, after developing repetitive strain injury in both arms in '97, He was an early user. Even with the very slow processors of the time, Dragon NaturallySpeaking was like magic. It would score close to 95% accuracy. I didn't know at the time that I would be a user of their products for the next 30 years. But what I immediately knew was that the processing power of computers improved. I could have access to much more powerful voice recognition software, enabling me to carry out my work despite my continued inability to type. By 2000, everything was going as I and others had predicted. The software had gotten better, and with better computer chips, accuracy and functionality were improving at a fast pace. What nobody could predict was that in the next 2 decades, despite advances in the underlying technology, consumers would see little of the gains thanks to a series of corporate disasters. Dragon was acquired in 2000 by a Belgian company, LearnOut and Housepee, that went bankrupt within months of the purchase, with its co-founders later Being found guilty of corporate fraud. You hate to see that.

Naughty, naughty.

So Dragon was then purchased by ScanSoft, rebranded as Nuance. I actually worked for a startup in college that got acquired by Nuance. Very, very talented team over there, which progressively deprioritized the consumer in favor of medical and enterprise markets. If you walk into a hospital today, you will see a lot of doctors and nurses dictating notes even before the Whisper revolution that just recently happened. Microsoft acquired Nuance in 2022 and continued the same model. The result, voice recognition software on PCs today is only marginally better than what was available in 2000, far less progress than the underlying technology would have led me to expect. This disappointing history lesson is one of the reasons why even greater curation is warranted when making predictions about the fast diffusion of AI models today. Breakthroughs in the infrastructure of a new technology need to go hand in hand with the right kinds of applications built on this infrastructure. And then, of course, you need the right marketing structure to reach potential customer base as well as inputs from consumers to act as the impetus for modifications and improvements of the applications. This is what we are not seeing for AI. This is what was completely bungled for voice recognition. What do you think?

A whole lot of mumbo jumbo, John.

Okay.

Let's move over to 60 Minutes because Darren Acevedo Mogul takes the other side of his own argument, arguing against himself with Brendan Foody from Merkur. Job category on LinkedIn. And just as machinists once replaced artisans, Brendan Foody insists he's creating a career of the future. Out of work?

Fear not.

Come train AI.

Literally all of the top 20 economists in America, they would say the exact same thing, which is that there's going to be more jobs in 10 years than there are today.

You're a top economist in America.

What say you?

No.

One of the greatest edits ever. That could happen.

What we are living through with AI is unprecedented. The first phase of the Industrial Revolution.

You're positive this is the same guy?

Yes.

In this article?

Yeah.

Wait, I don't think these are like— these two things are incompatible.

Break it down for us.

So the article—

yeah, because he is saying 5%, 5% of the things humans do are at risk.

No, he's saying that right now, if we keep building AI the same way, it will replace workers, but we need to build it a different way that, like, empowers workers. That's what the headline of the article is, Will AI Replace Workers? Not If We Build It Right.

Okay.

He's saying, like, right now, on the course that we're doing right now, it's going to replace workers.

Got it.

So, we need to shift courses.

Okay.

Right?

That's what I—

And his argument from an economic perspective is, like, that will actually increase productivity. That will be better, not just because Higher employment is good, but also higher economic growth is good. So we should be trying to solve problems that cannot be solved by people right now with AI, as opposed to just—

Just displacing. Yeah, just moving stuff around versus doing net new things.

Okay, I like that.

I believe that's what he's saying.

You're a top economist. What do you say?

No.

One of the greatest.

And here's the thing.

Yeah.

Top economists have never been wrong on anything. So, you know.

It does feel like it's a critical moment. There's been so many people building up to like 2027 is going to be the crazy year. 2026 is going to be the last normal year. Like, like something has to happen in 2027 that will answer a lot of the questions that have been being asked over the last 5 years, you know, since GPT-3 basically, which is like RSI, superintelligence, serious Serious move on GDP, serious like showing up in the statistics somewhere other than just debt. But I mean, at the same time, you see like weird things happening in in financial markets. Like gold is spiking, and like it's hard to put that on AI. But there's still like like there's there there's pockets of like okay, there's something odd happening that you might in retrospect attribute to the AI boom. Anyway, what do you think of Jeremy's take? What agriculture did to our bodies. AI will do to our brains? Make us stronger and taller? I think he's being negative.

I think he's being negative. Didn't it in some way make us like, you know, weaker because we're, you know, consuming more grains versus, you know? Yeah.

I mean, some agriculture made us a lot taller and stronger, but then some made us fatter for sure.

Yeah.

Because we overconsumed. And so, yeah, if you're overconsuming information, it's brain rot. Yeah, but a good encapsulation of the thesis. Anyway, where do you want to go next, Jordi?

We got to talk about Warren Buffett because he's just like you. He's just like you, folks. A viewer's guide to Warren Buffett's nightly trips down the YouTube rabbit hole. Glenn Close clips, 1950s commercials, Uzbekistan's version of America's Got Talent. The legendary investor can't resist the algorithm either. Feeling guilty about spending hours each day watching YouTube? Don't be. Warren Buffett is right there with you. Glenn Close clips, commercials from the 1950s, Paul Anka videos. The famed investor can't get enough. Buffett enjoys speeches by complete unknowns. He's a fan of Berkshire Hathaway's own content, too proud of how it performs on the video platform. He watches it all nearly every evening, usually from a recliner in his Omaha home. Warren's range is broad, says Sharon Osberg, a close friend and frequent viewing companion. It can be the silliest things. Buffett also goes down random rabbit holes, just like us. The next day, he'll come into his office gushing over a moving amateur rendition of My Way. In August, Osberg visited Buffett's home in Omaha. They spent hours on YouTube grinding. We watched a really good documentary on Roy Cohn, she says. And the Uzbekistan version of America's Got Talent, that would be Uzbekistan Got Talent. It's unclear. Easy to remember. Buffett, who recently turned 96 and stepped down as Berkshire's chairman, declined to comment on his YouTube habits. What's up with Sharon just blowing up, blowing up his whole routine? If I invite a friend to my home, I'll go dark. If I invite a friend to my home to watch YouTube videos with me for hours and hours and hours and go down the YouTube rabbit hole, And then, a week later, you're calling up the Journal, spilling the truth about my video consumption, my rabbit hole habits. I feel like Buffett had to be like, I want you to tell the Journal.

You think it's astroturfed?

Yeah.

I think, I want you to take this. You think he's paying people to do this stuff? I think he's as locked in as ever. I don't think he's doing anything for entertainment. He wants the other allocators to get comfortable. You know, I can spend a few hours on YouTube every night. Like, meanwhile, he's studying the blade every single night.

Yeah.

All of this given the benefit of 4D chess.

Yes. Occam's 4D razor.

Yeah, that is Occam's 4D razor.

Yeah.

This is what he's doing. Makes sense. What do you think of the new passport? People were going back and forth on it. Paul McGregor was saying, if I was redesigning the COVID of what is arguably the country's most important public document, I'd first hire both the best typographer and crest designer money can buy, then fiddle about with the layouts, probably. And Joe Gebbia fires back, hi, when you're done finding— when you're done fiddling, please send me resume of top crest designer. Make sure they can do better than the 1782 official national anthem. It looks like a very—

National emblem.

National emblem. Yeah, the redesign looks like a very tasteful and moderate iteration. Like I think people are reacting to this like it's like now it's all gold and it has like a bear and an eagle fighting on it and it features like the UFC octagon, which like was sort of where people's expectations are for these types of things. But this feels like surprising.

There's no UFC branding on the passport.

Yeah, no ads is a big step backwards.

Yeah, I will say that the passport text where it's kind of sans serif All caps and there's like spacing. That does look like AI slop, kind of.

It does.

This is like very common when you just have—

Too much— too wide of kerning is your problem. Okay. The kerning needs some work.

Yeah.

It does read a little odd.

That's pretty high-stakes graphic design.

It is.

Most stuff, you're like, yeah, we're going to make this, you know, pamphlet for a conference and nobody's going to read it. If they do, they'll look at it for a second. The passports, like we're going to make you know 100 million of these, and then they're going to carry them for a decade. Yeah. Representing the country globally.

Yeah.

They will be inspected many times.

I'm not. I'm not into passports. I don't. I don't want to leave America. It's like not a good thing. They can just get rid of them. Just stay here.

What else we got?

Oh, les chatons fat.

It's time.

Mistral.

Le Chonk.

They're calling it Le Chonk.

Le Chonk. Yes. They're leaning into the jokes. Mistral AI, the French Neo Lab.

Is it fair to call them a Neo Lab or are they just a lab?

They're just a lab, right? They're pre-Neo Lab.

Yeah. Yeah, I would say so. They're kind of like—

We call them pre-Neo Lab specifically?

Let me actually consult our nomenclature.

Okay. Oh, I think they were like National Lab or something. Something like that.

Always.

Up there with Lawrence Livermore. But they launched Mistral Large 4, aka LeChonk. 1 trillion parameters, natively made it multimodal, 49 billion active parameters. It is the best open weight model from US or Europe on aggregated benchmarks, state-of-the-art on critical workloads including cyber defense, manufacturing, and finance, and it surpasses closed frontier models on visual grounding, forged in Europe, end-to-end deployable from Europe via our own Mistral cloud infrastructure available to all via API today, working with cybersecurity partners privately. Open weights release end of October. So presumably they'll release the open weights for like the cyber nerfed version, I would imagine, something like that. I don't know. Anyway, what is your review, Tyler? Do you have any other data to share about Mistral's progress?

Yeah, I mean, it looks very good. I think it's very much on par with the frontier of the open source, so like the Chinese labs. It's very much a contender. It's not necessarily destroying all of them on every benchmark, but it's certainly, if people are worried about using Chinese open source models because maybe downstream, there's going to be some legal ramification or whatever. This is a very good choice.

It really does change the open source, closed source debate because—

Yeah. I think Mistral and Reflection, which is the American open source lab, we're now starting to see, okay, these models are actually quite competent.

Reflection came out with, it's called Beam yesterday. Whoa, not tapped in. You're like, sorry, I can't hear you. I can't hear you.

It's called Beam, yes.

Anyways, we can come back to that. But great to see them getting a new model out.

And Leo has the funniest benchmark of them. Says, it's pretty funny that it seems the one thing Mistral Large 4 is actually leading on is a benchmark about regulation. Another EU banger. Harvey, Harvey's legal agent benchmark is up to 15%, whereas other models are down at between 5% and 13%. Interesting. I wonder if that's intentional. I wonder if there's actually anything to read into that. I really have no idea, but it's certainly valuable. I'm sure Harvey and all the other legal companies will be excited because we, I mean, we read the whole thing about Harvey. Their margins being in a really rough spot for a minute when they were overly reliant on the most expensive frontier models. Now there's a ton of price competition between OpenAI and Anthropic, and then obviously there's a bunch of pressure from open source models. So it's truly the create the benchmark that you want the labs to hill climb, and they just created the benchmark for their exact workload, and then they're able to Sushi Commander says, what are you talking about?

Both Beam and Mistral Large are a generation behind the latest Chinese leading models. So putting you in the truth zone, Tyler.

I mean, you can look at the benchmarks. It's like, yeah, it's not leading them. It's like among one of the contenders. It's not the frontier open source model for sure.

Yeah, it's like in line, but a lot of the most recent round of Chinese open source happened like a couple months ago, which is always rough. 'Cause like you catch up to the frontier and then it's like, well, you're gonna take another 6 months to get something out where they're gonna only take another 2 months to get something out. So that window, but there's so many other geopolitical implications that lead companies to go one way or another. So we'll see, we'll see. Olivia Moore has some data on what's happening in the battle for consumer AI dominance between ChatGPT, Claude, and Gemini. A16Z pulled the data across web and mobile plus consumer revenue. ChatGPT is at 2x Gemini scale and 6x scale on web and 2.5x Gemini scale and 14x Claude scale on mobile. Interesting. I mean, I guess that makes sense. More consumer focused. You're pulling there.

The main thing here is just look at, look at the run from Claude. Yeah.

I mean, we saw this in the App Store, right? Like you would open up the App Store and Claude was at the top for a while and then and then fell down. And now ChatGPT is back on top and they're just going back and forth. It's the, the, the war for the frontier is just trading shots back and forth, back and forth.

Yeah.

Right now, Muse still at number 1. Vinted, who we're working on getting on, is number 2.

Vinted is still number 2.

So dominant.

The folks over at Vinted are, are really doing a fantastic job. Thank you anyways for tuning into TVPN today.

It's been an honor.

Fantastic show. We'll see you tomorrow at 11 AM Pacific.

I can't wait.

Leave us five stars on Apple Podcasts and Spotify. Sign up for our newsletter at tvpn.com, and we will see you tomorrow. Have a wonderful goodbye.