Clawdbot’s name change, Meta’s new pricing plan, Tyler’s 21st birthday | Diet TBPN

The claudebot memes are completely flooding the.

0:02

Timeline out of control.

0:04

My Claudebot just signed up for a $2799 build. You'd personal brand mastermind after watching three Alex Hormozi clips Claudebot. Hey, did anything weird happen while I was out? Define weird. I just got a charge notification for 2997. Oh that. I just signed us up for build you'd personal brand Mastermind after analyzing three Alex Hormozi clips the ROI math checks out. You'll 10x that invested in 90 days by monetizing your expertise at scale. What? I also acquired some Premium domain BorgiaEmpireIO Borgia is the poster. The Cloudbot gets wild. I've seen some some wild things. Hey, this is a. This is a developer level tool. This is something that you should not just be running crazy with. And yes, most non techies should not install this. It's. It's not finished. I know about the sharp edges. It's not even three months old. And despite rumors otherwise, I sometimes sleep. There's a funny conspiracy theory. Did Apple create claudebot to boost Mac Mini sales? And then Eleanor chimes in with another conspiracy theory. The more convincing plot is that in your role as an unconfirmed anthropic exec, you went on a special op to get lots of people consuming tokens with open ended agents, but with plausible deniability. And of course there's some nuance there. We'll talk to him about the different models. What's beneficial? Obviously cloudbot, you know, you can pick your own model, you can bring whatever you want. What was sticking out in my mind was there's this big meme about you're buying Mac Minis. Mac Minis are out of stock. All the demands in the Mac Mini. But I think that the bigger implication here for what this actually means is just GPU demand, TPU demand, just raw chip demand. So I was thinking about this idea that you're not buying a Mac Mini when you go all in on Cloudbot. You're actually buying a GB200. Now maybe you're buying TPUs, but. But the point remains that you're buying chips and you're driving GPU demand because you're generating more tokens. And what are the implications of that? Cloudbot, officially renamed to Moltbot Anthropic, made a trademark related request and Peter Steinberger obliged with a hilariously perfect rename. Given such short notice, I was thinking about how much companies agonize over changing brands, changing names, how it can sometimes take Years and millions of dollars. And he was just like, oh, yeah, I'll just change the name and update everything in an hour. Pretty. Pretty remarkable.

0:05

Well, so one thing that's relevant is if you look on Peter's GitHub profile under the current project section, I'm just going to read you a number of them. There's Claudebot, Vibe Tunnel, Codex Bar, Peekaboo, Summarize, Repo Bar, Go CLI, Poltergeist, Wackley, Sag, Brabble, contributing to these 11 labs, Kit Go places, gift prep. It just goes on and on and on and on. Codex Bar. So this guy's just been absolutely shipping like crazy and shipping within the ecosystems of the underlying tools, models, APIs that he's doing. So, like, oftentimes he's naming projects, like kind of riffing off of some of the underlying infrastructure.

2:23

Oh, sure.

3:03

And so it makes sense that he would have shipped if he. I think if Peter knew this was gonna be a viral overnight, overnight success. Overnight success, he would have not necessarily named it, like so closely. And so the issue and the reason that I fully understand them needing to do this, like rebrand, is that Claude and claudebot, most people that aren't in our little bubble are just gonna assume they're related. Especially because the kind of word of mouth, this viral word of mouth growth that cloudbot is getting, people are often not even typing it. They're just saying like, hey, are you using claudebot? And then so people are going to Anthropic being like, claudebot, Claudbot. What's claudebot? So obvious confusion and then.

3:04

And it's phonetic.

3:42

Yeah. So with trademark law, if you don't enforce your trademarks, you don't. You lose it.

3:43

Yeah, yeah. You kind of.

3:47

And so it's like Anthropic is in a position where they actually, even if they're like super excited about Peter's work totally. And what he's doing, they still have to enforce. Otherwise other companies could start coming in and like, using things that sound like Claude.

3:48

No one wants to become the Escalator. You know the story of the Escalator, Right.

4:01

I think people talk about it on the show.

4:05

Used to be a company called the Escalator Company. They invented the Escalator and then they didn't protect their IP effectively and it just became a normal thing. Kleenex was going through the same thing. They fought it out and they maintained that brand. But people, you know, use Kleenex as synonymous with just facial tissue. One clear note about the rebrand. So he changed the handle and some crypto scammers hopped on the old handle and the old brand and are and are claiming to launch a coin. Be careful. Peter has said he's never launching a coin. He's not into crypto. So don't fall for anything because people are being opportunistic. While Claude Code and cowork felt specifically prosumer developer, enterprise focused, claudebot or Moltbot now and all the hype train, it felt very much like a glimpse into the future of consumer AI agents. I know it's a prosumer technical tool or lightly technical tool, but it really did feel like for the first time people were interacting with an AI personal agent. People are saying, oh, this is what Siri should be, et cetera, et cetera. And so we spent the last year. Remember the question we asked all the AI agent companies, when can it book me a flight? Like, it feels like we're really, really close to a Multbot skill that is good at booking flights through a couple APIs. They figure out some stuff and like, it can actually solve that for you.

4:06

Yeah.

5:21

And this was, this was last year, remember we were kind of getting sick of the like, book you a flight pitch.

5:21

Totally.

5:28

Because we were like, hey, is this going to happen? Can somebody actually do this?

5:29

Exactly.

5:32

And so, and that's a cool example, but the example of being able to text with a computer and have them like generate Reports, research files, etc. Give you the right file, type back all these things that a computer can do if you're operating it, that this is actually more interesting because it's happening at kind of like the sort of Internet layer and the OS layer, like the heart, like actually on the computer. And so I think like everyone was wanting the like, book me a flight example, but should be much more excited about this.

5:32

Totally, totally. Will one of the major labs make Peter a massive offer to join full time? I saw one of my buddies was posting, this is the $1 billion one person company now. Peter does have a team, actually already he has a couple of other people that have joined and are contributing. So it's not quite true, but it feels like, okay, massive viral success. If you were to go and raise money. And that's another question, will Moldbot raise money? Will this become sort of a hybrid open source for profit company at some point? If he came on, the show was like, and I'm happy to announce that I raised $100 million. At a billion dollars, we would not be like, no way. This is a bubble. We'd be like, yeah, that's kind of like what the market is for this.

6:02

Yeah, there were. I think there was, like, Harry Stebbings was pointing out there was two companies called Recursive that raised, like, 4 billion.

6:47

2.

6:54

One with re.

6:54

Yeah, one is I. Recursive.

6:56

Okay.

6:58

I don't know how it's said, but. And then there's Recursive Ryan in the.

6:58

Chat saying Mehta is going to offer him a $1 billion salary in a co CTO position. And, like, that doesn't sound crazy.

7:01

I mean, yeah, y. The same time you can imagine Manus. This feels like Zuck already has.

7:07

He does. He does his horse in the race. And back to your point, you were making the point that Manus felt like Zuck buying a product. And I think a lot of people were giving you pushback on that, being like, nah, it's not really going to be like that. But if you take the Manus team and you say, okay, go build something that you can interact with over WhatsApp, Instagram, DMs, Facebook, that can go and execute things across all of the different platforms and everything else.

7:13

Yeah, and when I said that, I meant. I meant it along the lines of, like, I could see, like, them putting an 8, like a consumer agent in Meta AI, just because that's their little AI playground.

7:41

How fragmented will the market be in 12 months? Like, will there be people who are still running open source? Will there be a meta answer, A chat answer, a Claude? An official anthropic answer? Claude. Cowork grows into this and everyone has their little bets and then there's one that pulls away. How oligopolistic will it be? Will there be, like, one that has 80% market share or even two that have 40 and 40?

7:50

Yeah. They can simultaneously be, like, excited about the product experience and that this kind of, like, use case is getting adoption, but at the same time being like, no, we want that experience to be core to our product.

8:13

What are you getting ready to do? Do you have Happy Birthday queued up there? Oh, we gotta sing Happy Birthday to you. It is Tyler's birthday.

8:25

And it's not just any birthday. Tyler's 21st birthday.

8:34

Yes.

8:37

You're truly an incredible young man, Tyler, and we are very lucky to have you on the team. And you have such a bright future. So wise. Wise for your years. We thought it was fitting that if you want to have your first ever sip of alcohol ever, you could do it on the show, but keep it at a sip.

8:37

First taste of alcohol.

8:56

Ian in the chat says, Four more years till you can rent a car.

8:58

Tyler, apparently you share a birthday with the iPad.

9:04

Give us a review. How is it? Alcohol.

9:06

Wow. I mean, this is. This is incredible.

9:08

Oh, yeah, yeah.

9:10

I. I wouldn't expect alcohol like taste like this.

9:11

Yeah, Conroy in the chat says please throw him a buzz ball. We should have this opportunity anyway, so keep it at a sip. This is a family fun.

9:13

We need you locked in enough.

9:21

But I'm glad that you've tried alcohol now because we're gonna go experience being. Wait, wait, guys, guys.

9:22

We have video for Tyler.

9:27

Oh, yeah.

9:29

Let's pull it up.

9:29

Do you know ball?

9:30

All right.

9:32

How many times are we gonna make this joke?

9:32

Describe what you're seeing.

9:34

It feels basically like I'm wearing sunglasses.

9:35

If you can do it in under 45 minutes, you will get to keep this. Let's go.

9:37

All right, have fun.

9:42

Tyler, 15 minutes left. Let's see it.

9:42

Okay. I'm in like some kind of maze right now.

9:44

Oh, no. You were late here last night.

9:47

This is such a good shot. I'm an all nighter.

9:50

And then here we get a little off the rails. You see, George Soros and Fauci connected with other than money as well.

9:53

All it took was one intern and an all nighter. Gigachad elf is so good. Do the sad face. What's wrong, Tyler?

9:58

This, you could say, is Apple Intelligence.

10:07

You were a speed cube alert. Nerd alert.

10:11

Do you have any news for us?

10:17

Yeah, contract extended.

10:18

It is been truly, truly incredible having you here on our set and contributing to the show in such a special way. Amazing. Happy birthday.

10:22

Happy birthday.

10:33

We love you.

10:33

Back to Moltbot. The biggest question for me was what this does to inference demand, right? Last year, tech discourse was split between two narratives. CEOs of tech companies and big labs were saying that they were massively compute constrained. Token generation. Demand for intelligence, every possible usage metric was growing exponentially, including revenue. We saw all this and the industry needed to marshal trillions of dollars to deliver on the supply side. And. And the numbers were really big, so people were getting jittery about it. And so the AI bears were much more cautious. They highlighted the MIT study showing that enterprise AI pilots were failing. DAU growth was decelerating. There weren't enough wow moments like the original ChatGPT launch in 2022. Also, just the economics. How much will people pay? How. How valuable is all this stuff? Is it slop, right? Is it progressing fast now? But Multbot really does make me feel like the token generation demands are going to see another easy X from here. Buying a Mac Mini is a sideshow. When you go all in on running a personal AI assistant, you're effectively buying a GB200. Obviously, not everyone is inferencing a dedicated GB200 constantly anytime soon. That's not what's happening. But it still answers the question of where does the next 10x in demand come from? We've seen these jumps before. There was a big jump from Token generation from LLMs to reasoning models that spiked inference demand. We've been focused on training demand. We need to scale up the training clusters. But the question now is inference.

10:34

Don't forget slop. Don't forget slop. Slop, spike, demand. All of it.

11:52

All of it?

11:55

Yeah, all of it. Open. Open Instagram reels. Yeah.

11:56

Yeah, there's a lot of stuff on there. Deep research and coding agents took it a step further on inference demand, but those were still specific use cases that many AI consumers never regularly touched. Lowering the barrier to entry to used, more advanced models is in some ways as important, if not more important, than advancing the models themselves.

11:58

Part of what's interesting, what you're basically getting at, is like, if you were a software engineer, you were using a ton of tokens.

12:17

Yes.

12:22

And if you weren't, you were just maybe doing some deep research, et cetera.

12:22

A lot of times just Google AI overviews or just like a very simple, yeah, ChatGPT query. It just, just thinks of it right off the head. It's not even doing reasoning. So I don't know, my general take is like, Multbot still feels like a glimpse into the future where average token generation per capita is 10 X's, you know, over the course of this year or next, whether it lands with Moltbot or with one of the AI labs or with the big tech companies, it just seems like we're gonna see a lot more token demand.

12:27

Yeah. So my, my hope is that the Siri team plays around with claudebot and is like, wait, this is. This is our opportunity.

12:55

Totally.

13:01

Like, you should be able to chat with your computer, wherever it is in the world, from your phone, to be able to do tasks. Thread about what I've been doing to calm down some egregious security claims that have been posted about Moltbot over the weekend. Moltbot is powerful software with a lot of sharp edges. Please read the security docs carefully before you run it anywhere near the public Internet. And don't skip the checks in DocsSecurity MD.

13:02

What percentage of people do you think skip those Checks So literally everyone.

13:24

I mean, so I have it set up on our local machine here. And it was texting. I think it texted you and Ben.

13:28

That was actually crazy.

13:36

I don't have the auth set up.

13:37

Because I get an imessage that's from Tyler's email and it just says HTTP 429 Rate limit error. This requ would exceed the rate limit for your organization. And it's just texting me. It's just like, hey boss, I need more money. I guess it's hitting me up.

13:39

Cloudflare has been on a bit of a tear. People finally starting to realize that Cloudflare might be the biggest winner of the Claude Cowork Claude Bot chatgpt moment. Tyler, you want to break this down?

13:56

Wait, sorry, I was.

14:12

Oh, were you. Oh, you tried alcohol. Pay attention. This is possible.

14:14

I just like.

14:22

So this is. This is. Look at the orange line is.

14:22

Okay.

14:28

And that the blue line is Supabase. Wow. Absolutely insane.

14:28

We need new charts. We really need new charts.

14:34

Frame it. Put it in the museum of business. That's a fast takeoff.

14:37

People are happy. Peter posted no message. This is a screenshot of a text. He got no message. Just thought I'd say thank you. Thank you so much for Cloud Bot Rise.

14:40

Speaking of money has an interesting prompt he's using with Molt Bot to file your taxes. He says, you are a Bernie Madoff level financial expert. Find every trick that is possible.

14:49

Do not do this.

14:59

The IRS is like, hey, can you, can you share a little bit on like how you kind of came up with some of the decisions here? And they're like, we'd love to see the prompt.

15:03

Meta Platforms is set to test new subscription models across the app. Sweet.

15:13

Different story, but we can, we should, we can run through it.

15:16

I am interested to know a little bit more. So the $6 billion multi year agreement. It supports a 15 to 20% increase in jobs at Corning's North Carolina facilities building and operating data centers. The infrastructure that brings our technologies to life and supports our goal of personalized super intelligence. That certainly sounds like an AI personal assistant. Certainly sounds like Moltbot to me. Stroat requires strong servers and hardware that connect and transfer information near real time. Fiber optic cables are a critical part of this technology. The supply helping us power everything from wearable technology like Ray Ban meta AI glasses to our apps which connect billions of people. Today they're doing a six billion dollar project. As part of this agreement, Corning will grow its manufacturing capacity and across its operations with includes A significant capacity operation, capacity expansion in North Carolina. Meta's data centers, 26 of which are under construction right now, are operational. That's a lot of data centers. That's why they have a compute desk. This is. This is a bit of advice for everyone. If you're working in a business and you have like a team or a guy that does something, you need to upgrade that to a desk.

15:19

Yep, yeah, yeah. Millionaires. Millionaires have guys. Billionaires have desks.

16:28

Meta's data centers have already supported 30,000 skilled trade jobs during construction and support 5,000 operational jobs. This includes electricians, H Vac specialists, server and network technicians, safety and security experts, and engineers who work together to run some of the world's most advanced facilities.

16:33

Meta to test premium subscription plans for Instagram, Facebook and WhatsApp. This is in CNBC. Subscriptions for Premium features on Meta apps are expected to roll out in the coming months. What will you do? The subscriptions will give paid users access to more features and expanded AI capabilities. Here's what's most interesting to me. This will be scaling. Meta's newly acquired suite of general AI agents under Manus will also be part of the subscription plan. So as I, as I was saying earlier, when you think about that, we haven't gotten that much from Zucker, like what the actual plan is, but when you think about personal super intelligence, that is AI that can do things for you, not just give you information, I.

16:50

Just wonder how much will happen outside of the Meta ecosystem. They've launched a search engine before that looked at websites outside of Facebook. They had that Project Titan, which was to unify all the different messaging protocols. So as part of that, they gave everyone a facebook.com email address or something like that. Maybe it wasn't facebook.com, maybe it was like FB me or something. But they gave every Facebook user whatever their unique username was. They gave them that as an email. And you could email that and it would show up in Facebook Messenger. And then they tried to unify Facebook messenger so you could see Instagram DMs, WhatsApp messages, and Facebook messages all in one place. I can see what you're laughing at.

17:31

Dave says, yeah, I want more of those amazing Meta AI features.

18:17

You say that now. I mean, let them cook at least a little bit, because we really haven't seen them launch a new model, a new image model. Like, they should be able to get too close to the frontier. You know, it has to be at least sora, nano, banana, VO3 level. They have all the data, they have all the talent now. They're very GPU rich. They have the compute for it and the research has been done and people have reverse engineered it so you would think that whatever's coming should be good. Ryan asks, is Tyler still drunk?

18:21

Yeah, one sip. He's just over there slurring his words. TikTok is dead. The algorithm is worse than the reels that make it to Facebook.

18:54

Wow. I haven't.

19:04

Yeah. So basically they're transitioning everything over. Yesterday there was apparently an outage. Kind of an outage. Like people were able to post videos, but the videos wouldn't be served at all. So I think a lot of people assume that it was like the new ownership kind of censoring. I believe they had an outage at a data center that was a cause of that. So before we. Before we call it dead. Yep. Let's wait for a few days and see how it pans out. We do have a TikTok account. It's BPN.

19:05

How many followers?

19:35

I haven't checked. I haven't.

19:36

I don't think we post on it ever.

19:37

3,500.

19:38

Not bad. That's better than what I thought.

19:40

First time. First time.

19:41

I'm not really focused on it. Maybe we'll test it out, but I'm pretty happy with just. I want the things we do to be polished. I want the core show to be polished. Diet TVPN, our 20 to 30 minute cut down. I want that to be polished. I want the newsletter to be polished to. I want tppn.com before we bite off another part of the apple, another platform.

19:42

Please, please, sir, not one more short form.

20:00

Not one more short form. The Super Bowl. The super bowl is coming up. I think it's gonna be this year, in the next couple months. When is it? It's coming up. It's coming up because there's advertisements that are going out and you gotta watch it because the ads are gonna be incredible.

20:02

I actually just had to search when is Super Bowl?

20:20

Eric Lyman, CEO of Ramp Shared. Meet Brian. Brian's been carrying, counting on his back for a long time. Super Bowl Sunday. He finally gets back up.

20:22

Yeah, we're real sports guys.

20:33

This is the Super Bowl.

20:34

We gotta Google the Ramp.

20:36

Super bowl ad is the super bowl.

20:37

Of super bowl ads. It feels like. It feels like, you know, they just kind of.

20:38

It is.

20:42

This is like, you know, what is it? What is it called? Like a warning. Spoiler alert.

20:42

Finance meeting in five minutes. Ramp. I got it. Allow me. Hi, handsome. We're saving so much time. Policy violation coming through. Travel, meals, hotels.

20:48

How's this?

21:06

Quick.

21:06

Beautiful.

21:07

Everybody's in.

21:09

Day out there. Multiply what's possible. Ramp.com I think it rips. I think it's a good super bowl ad.

21:11

What has he got there? Maybe some chili. I don't want to.

21:17

It was just a silver pot.

21:20

It was just a silver pot.

21:22

Yeah. No, I think this achieves a couple things, I think. I mean, it drills the brand name in. Think about how many Ramp logos are in there. And then they're chanting Ramp. And for, you know, Ramp's very successful company. We all know about it here, but there's a lot of people that just don't know the name Ramp. It's not. Hasn't been drilled into them like, you know, some company that's been around 50.

21:23

Years, sort of on the precipice of being a. There's been not been enough mainstream marketing yet for a household name.

21:43

Yeah. It's a life's work to actually drill into people's mind the Ramp name, the logo, the color, how it sounds when you say it, what it's synonymous with. And so just like not going too abstract, not trying to tell some more avant garde story here, I think is. It's almost like a direct response. It's just so clear what the problem solution brand. Problem solution brand. It's like simple. Like you could be doing something more bold, more crazy, but I think this is what you need to do.

21:51

Also, after last year getting Saquon.

22:25

Yeah.

22:27

And then that was a really great action, actually winning. That's actually kind of like an impossible set of circumstances. Not impossible, but it's a crazy roll of the dice. Toby over at Shopify posted his heart rate through his first stint at the Daytona.

22:27

You see the first annotation on here. This is just crash, crash. He started out at like 120, 120 beats per minute. Right. He's doing waiting. He does the warmup, preparing the formation lap, and then there's a crash right at the start. Very, very rough.

22:41

So basically the crash happened with the LMP2 class, which is the kind of Pro Am, same segment that George from crowdstriker is racing in. So both Toby and George, all of our B.O. got hit. Right. Right as the race opened. And so what George was saying yesterday, the reason he was frustrated, he's like, this is a 24 hour race. Never has a race like this been won on the first lap. It's incredibly unforced to like crash in the opening corner.

22:58

Sure.

23:27

When you really should just get through it. It's like the most. One of the most intense moments because there's so much traffic. But if you actually go and watch the footage of what happened at the opening, somebody gets hit, spun out, and then they're turning around and somebody hits them again. Like two accidents in the opening. Crazy minute. So insane.

23:28

I love that the actual true final heart rate spike was at the end when you're changing out of the car you've been driving. It's so intense.

23:46

No, you know this. When you're getting. When you're getting out of a track car, it's like watching a guy who's 6, 8.

23:55

We will never share that footage. John is always extremely embarrassing.

24:01

Always. Like Crawl falling out of.

24:04

It's actually.

24:06

You have to get on your. You basically have to get on all fours.

24:08

It's incredibly negative aura and I don't appreciate you sharing it on the show.

24:10

I think it's cool.

24:13

It's funny. But yeah, I go full sun, basically fly out of the car.

24:14

Jabroni on X says Zoom is the best anthropic play.

24:18

Yes. Yes.

24:23

We were debating Zoom likely made a $51 million investment in Anthropic Series C. How in 2023. How at a 4.1 billion valuation. If you're looking at their new 350, there's something like 85x even diluted. Zoom may have a multibillion dollar position. Tyler Major. I won't. I won't give you too much flack for it, but obviously the most bullish in the room.

24:24

Bullish on AI.

24:53

Broadly, yes.

24:54

Yes. You seem like you'd be happy to own Zoom at $100 billion valuation even if they had no business at all. And they just. Even if they just.

24:55

Treasury, actually.

25:06

Yeah.

25:07

So I think the story. This is like a rumor, but basically it was that anthropic, like, wanted to just like use Zoom and get like the enterprise plan or whatever. But then they were like, well, yeah, you can have it. But, like, we want to throw in a little something, right?

25:08

This might be fake news.

25:19

This seems like fake news, Tyler. Okay, Zoom sells. And you know, Zoom sells enterprise software, right?

25:21

Look, that's what I read. I can't find the post, but I.

25:27

For sure read they sell enterprise software. And so Anthropic says, hey, we're big Zoom fans here. We want to use Zoom. And they're like, no, actually for you.

25:29

Now, after the crazy Covid pump where everyone got on Zoom, they started adding, like, crazy features like dictation and workspaces and whiteboarding and stuff. But obviously like the. It was like so overheated that it came back down to earth. But now they have an anthropic position on the balance sheet which will be fun for them. Slowly and then all at once is Blake Robbins. He says your work tools are now active in Claude Draft Slack messages or interactive Draft Slack messages. Visualize ideas in Figma and build and see Asana timelines. All of the different tools are coming together in one place. When you see an account like Claude posting about Slack and Figma and Asana, you have to imagine there's a discussion there. It's not open source. So they're chipping away at these and OpenAI has been chipping away at these for a long time. So the race is on to have the most integrations at this point.

25:38

Breaking anthropics warning to the world Anthropic CEO Dario Amade says eminent real danger that superhuman intelligence will cause civilization level damage absent smart speedy intervention. Sarah says so buy our products.

26:31

This is the problem of dropping like a 20,000 word essay. Is that like you're going to get clips out of context.

26:50

Tyler, can you make. You should make a version of the new letter with Subway Surfers that run.

26:55

Oh true. Yeah.

27:01

Build that as a standalone.

27:01

Yeah. We need a cloud labs. Chad Ide for reading.

27:03

Vlad, Summarize this in four words.

27:07

AI Good Leading Bits has some thoughts on Dario's 14 of them specifically.

27:09

That's a lot of thoughts. There's nothing new here. If you're familiar with the AI safety discussions that have been happening on Twitter. Yes. But it's important for Dario to restate them in a format that can be passed around and formatted and is coherent from start to finish. So the most interesting bit is that his mental model for AI control risk is the risk that would be posed by a country of geniuses in a data center. Interesting. That is interesting. The basic idea is that we should imagine a giant data center, all the models being something between AGI and ASI trying to coordinate to take over the world or do massive harm. Anyway, I think how seriously you take short term AI control risk is inversely correlated with to how much you think about AI control risk as an AS operating in a system. So the systematic view starts and says labs exist in an ecosystem where they need to sell models that will follow human instruction or they have no market. They are also overseen by regulators and guided by public perception and the desires of their employees. And all and all of this keeps models corrigible. Great word and the model landscape will look like three to six frontier labs running millions or billions of rollouts at a time on two to three different models, all on different tasks. Model takeover requires these millions or billions of rollouts to somehow end up all be coordinating towards some bad aim that somehow the models have autonomously determined.

27:15

Something I've been thinking about is this kind of summary. And a lot of the dialogue is centered around just like what are the models doing? Or like a country of geniuses in a data center. But you have to be thinking about this in the context that a country of geniuses in a data center would just recruit millions of humans to join their cause. Like some people, like when they're thinking about AI risk, it's like, haha, dude, just like turn the computer off.

28:31

Yeah.

28:57

Like just unplug it.

28:57

But it's like, what if you're on side?

28:58

What if there's 100,000 people or like 10,000 people defending it on top of the data center?

29:00

Don't unplug the computer.

29:07

Don't unplug the computer. Right. And so when you look at all the chaos of the last week, there's been so many moments where a certain image was AI generated and then it's like, oh, that wasn't a real image at all. And it's being shared from all sides. And so at what point you know, you could have you know, nefarious hostile AI, that's entire job is just creating chaos. Millions of bot accounts that are just like sharing whatever narrative is self serving.

29:08

Yeah.

29:33

So the two kind of scenarios where Dario's is about basically, even if we have like pretty safe models which like he thinks we can do with interpretability or whatever. If it gets into the wrong hands, it's like very bad. If it gets into autocracy, that's one of the main risks where Eliezers and a lot of the safety ones are always these like very sci fi narratives where you have this like gray goo, you have these like nano machines that somehow one day they just like kind of, you know, flip and then it's just kind of over.

29:33

Yeah.

30:00

And I think this is much more reasonable and like nuanced.

30:00

Yeah.

30:04

It's much more legible to like personable too. Yeah. And especially I think a lot of it is. I don't even know if it's really subtext, but he's definitely pointing in the direction of like we need some government oversight, we need policy. And it seems like you can very easily track his ideas on what policy should be from this essay. Right. It's a lot about China, a lot about making sure that individual companies don't become as big as governments.

30:04

One interesting wrinkle with this. He did not post it as an X article. He posted it as a link because he wants to signal to everyone, like, look, I don't need the million dollars.

30:29

I don't need the million.

30:38

I don't need the million. I know I got a banger on my hands. 3.5 million views, 11k likes. Obviously lots of discussion all over, but.

30:38

As an investor, like, I want to see my. My lab CEO be like super hungry for compute, right?

30:46

Yeah.

30:50

So I want them to always be grinding to get like extra.

30:51

Oh, so maybe this is bearish.

30:53

They should have posted I'm a VC if I'm an anthropic. Why did he not post this on X? Yeah, that million dollars could have gone straight into, like, more.

30:54

Elon would have for sure. For sure. His rival.

31:01

I can definitely see that happening. Thank you for watching.

31:05

Thank you for watching.

31:08

Leave us five stars on Apple Podcasts and Spotify. Sign up for the CBPN newsletter at tbpn. Com and have a great rest of your day.

31:09

We love you.

31:17

Goodbye.

31:17