GPT-5 Hot Takes, Rahul Live from The Ultradome, Doug From SemiAnalysis, Timeline Reactions | Doug O’Laughlin, Rahul Sonwalkar, Mitchell Green, Ben Schleuniger, Merrill Lutsky

You're watching TVPN. Today is Friday, August 8, 2025. We are live from the TVPN Ultradome, the Temple of technology, the fortress of finance, the capital of capital. We are joined in person by Rahul Sunwalker. Did I say that correctly?

That's perfect.

And he is here because we are crowning him the king of the application layer. Never looked down on the future. First ballot hall of famer. They said, don't build a rapper. Don't build a rapper. You steamrolled, he didn't listen. And he's built a beautiful business at gp.

It's a good product, sir.

It's a good product.

When asked if value would accrue to the model layer or the application layer, why not both? It's a good product, sir.

Why not both? What was your reaction to GPT5? Is it going to make your life easier? Is it? It's not going to put you out of business. Right?

It's not putting us out of business. It's making our product better, basically making every AI application layer product better. Also, it's half the cost of O3, so it's much cheaper. So it helps you, helps your margins. It means you can.

Don't say that out loud though, because you don't want your customers to ask for a 50% discount, right?

Well, so we pass on the savings to our customers and what we do is we have the model generate more tokens, think for longer, and then produce better results.

Yeah, yeah.

Because we're still in the era of just let's get the best possible result. Let's get, let's just actually like the. I don't know, do you have a, do you have a rough benchmark of like, cost per task? Like, if I, if I want to, you know, crunch our analytics across, you know, look at the trends on our views on X, YouTube, Instagram. We have a bunch of data sources. Sometimes they're in spreadsheets, sometimes they can be linked. I export those all. I have a bunch of CSVs, maybe I put them in a database, I link it up to Julius, and then I want to do an analysis. That could be a couple hours of a data analysis time. That's going to be hundreds of dollars even at the low end, probably thousands of dollars for like a simple analysis just on a opportunity cost basis for an individual employee. How much are you thinking it should cost? For the modern frontier best model with the most thinking, how much should that cost on inference?

So there's a couple of ways to think about this. The way we Think about this is how much would it cost for you to have a data scientist or a data analyst for every one of your employees? Your operations team, your finance team, your marketing team, your product team.

It would pretty much bankrupt every company.

I don't think we can hear you through that.

All right, all right.

You still got a. The spacemen are out. We're not going to space today. Although Firefly did IPO up 36%. If you didn't see the news, very good news. Firefly stock surges 34% in debut. Congrats to everyone over there.

I love the, the physical newspaper.

We love physical newspaper.

You got to do that.

Yeah, we're maxing. We're maxing. We read the Wall Street Journal. Today is a special day. It's Friday, so it's the mansion section.

We're news maxing here.

How many pools do you have?

We're news having right now of zero. Zero right now.

Well, you got to. Because the new thing is having two pools, a pool for every season. People are increasingly getting both indoor and outdoor swimming pools. So. Yeah, get on Zillow.

Get on Zillow. Zillow. Maxing here.

Okay.

You were telling me how much. So, yeah, I mean, it seems like, you know, uh, most, most of the application layer will be, you know, productivity tools a la Slack, a la a, you know, like Adeo, our salesforce, our CRM partner or something where, you know, you're doing like seat based pricing almost. Maybe there's consumption based pricing, but you're, you're kind of distributing the cost. You're making everyone slightly more productive and you're charging, you know, on the order of tens or hundreds of dollars per employee per month, something like that, right?

Absolutely. I mean, it's not just slightly more productive, but it's also like getting insights when you need them.

Right.

Sunday night, you're prepping for a big meeting on Monday. You can't reach out to your data analyst and get your insights in that moment. And so the convenience of having an AI that can help you with that is just invaluable.

Yeah, it's going from 0x to 1x engineer all over the org. We've seen this with a lot of the Vibe coding tools, figma and adding like Vibe coding to that product. You've taken designers and you've given them the ability to write just like a little bit of code and that's really helpful. And you're doing that for data scientists. And not just data scientists, but actually business operations people who probably would be intimidated by an IPython notebook, presumably.

Exactly.

Nailed it. Okay, I want everyone's feedback on my take. Vittorio had this post. He said, sam Altman's doing the Apple stance. TM, it's over. And I think that the reaction to GPT5 yesterday was. Was. Was interesting because there's a lot of people that say, like, it's better model, it's cheaper, it's good, it solves, it moves the ball down the field. It's a good model, sir. But I think people were mostly reacting to. They had expectations of super intelligence. They had expectations of God in a box. There's been so much rhetoric around that. Like, you know, that step up from GPT3 to GPT4 was insane.

Like, five just felt like a big number and it felt like we'd be discovering and novel science.

Totally, Totally. Yeah. Everyone was expecting like a binary qualitative jump where, you know, everyone recognized that, you know, GPT, when Chatbot dropped, we passed the Turing test. And the next. The next hurdle is like, I don't know, maybe super intelligence, whatever that means. Like, you know, massive. You know, just, you know, you just hit it with a prompt. It just solves everything. It does everything. Every other startup, it kills all the rappers. Like, the expectations were just so high that it was hard to match. So even. Even though there were a bunch of solid improvements and remember, the number one thing that I was asking for was just like, get rid of the model picker. Like, and I had. I actually was playing around with GPT5 yesterday, and I was really happy that I was able to say, hey, think about this. And I didn't have to go to the model picker. And it just went and just kicked off a reasoning chain. It was great. Got me a great answer.

But.

But power users so far are very upset about this. They want the model picker back.

True.

True is what I've been seeing generally,

but that always happens with these consumer products. Like, I remember when anytime something would switch to an algorithmic feed, all the people that were like, no, I perfectly curated my list. This happened in YouTube. Back in the day, the default YouTube view used to just be your subscriptions. And so you would never see a video unless you subscribe to that person. Terrible for discovery. And. But all the hardcore YouTubers loved it. Because if I put a YouTube video out, I know that my audience is gonna see it. Now I gotta do.

You actually had distribution?

Yeah, it was more like a subtitle.

Earn it every single time.

Yeah, exactly. And. And the same thing happened. I remember there Were like protest groups on Facebook when they launched the newsfeed. It's like the most dominant like product of all time. It's like incredibly valuable protests right now

on Reddit, people that, oh yeah, I want the old 4404 or 5.

Yeah, I think, I think those voices will be like, personally, I think people will get over it pretty quickly and I don't think that that those particular. That, that that small cohort of like chattering. The chattering class will be, will be like, they'll get over it.

In the clanker economy is in charge.

What do you have for me, Tyler?

I don't know if I agree with that. Like there was that whole Funeral for Claude 3.

Do you see this? Oh yeah, yeah, yeah, yeah. I hear that in person. Right?

Yeah, it, I think some people like really like the personality of certain models.

Yeah.

And those are like, it's not just intelligence, it's like how it like talks to you.

Yeah.

And if people would make like some kind of connection with that.

Yeah, I think it's, you know, I mean how many people, how many people attended that funeral?

Yeah, I guess it was a party.

As a percentage of the hundred million DAUs of these apps, like where are we, like 1%? No, it was like 40 people probably. Right. Like, it's just, it's just not, it's just. I mean, yeah, there were protests at, at Facebook HQ when they rolled out. Like people went to Facebook hq, were like, bring back the old feed. And it's like, yeah, now we're two decades into the algorithmic newsfeed and it's the most dominant consumer social app. It prints money and most people really like it. And the revealed preference was like, it's good enough. So anyway, my will say I'm just

going to read through Reddit's reaction, please let's go over to legrate R ChatGPT. GPT5 is the biggest piece of garbage. Even as a paid user, the people are not liking it. Another one, OpenAI just pulled the biggest bait and switch in AI history and I'm done. Another, if you miss 4.0, speak up now. Contact OpenAI support. Deleted my subscription after 2 years.

This is like, contact your senator, call your senator, speak up now.

Crazy. I mean, how many people are in the Google subreddit, like, complaining about various changes to like, the Google algorithm?

GPT5 is clearly a cost saving exercise. They removed all their expensive capable models and replaced them with an auto router that defaults to cost optimization. That sounds bad. So they wrap it up as GPT5 and proclaim. It's incredible.

I mean, there's so many times when I fire off an O3 query that a 4.0 could. One shot. Like having a model router makes a ton of sense, even just for consumer experience, of like getting a. Getting the correct answer faster.

A lot of viral posts from people just canceling their subscriptions.

But how many, you know?

Well, I'm just, I mean, I'm just providing context.

I'm not saying you think ARR goes down next month. Well, no way.

Well, you know, one intent. How many miles does Chad GP have? 700 million. Something like one in seven. One in ten people in the world.

They have 100 million. You could back into this. And there's roughly 100 million people in the U.S. that use it weekly.

Yeah.

Based on that 700 million number and the percentage that are outside of the 85% of their weekly actives are outside

of the U.S. just like one in ten people in the world aren't. Clanker Mouse. So it's kind of, you know, they're thinking about the bigger market, I feel like, in some ways. And then it's like, you know, when you want to get to the one to like the remaining 90% of the users, do you want a model that thinks for longer? You know, you want more personality. So I think they definitely leaned in on personality.

Yeah.

Which I think is interesting. I like what Tyler said, you know, it's kind of different than Feed in some ways because you, you know, you have this like, person you talk to. It's like, you know, it's like a relationship and then it just like switches up on how it talks to you.

Yeah, that makes sense. Do you. Do you. Do you talk to LLMs?

I'm shy.

I don't really talk to them. I mean, I treat them like. Like something I delegate tasks to.

Yeah.

And I do that a lot. And I'm definitely in the DAU 30 minutes a day. Love ChatGPT, but my workflow is I dictate go pull all this data together, put together a report. And I don't mind that it's using a lot of bullet points. I don't mind that it's using a lot of tables. Like, I want that result. I want it to look like the result that I get from Google, but just more hydrating.

I do think it's interesting that a lot of people are reporting that they're hitting. They're getting rate limited within an hour of usage. As a pro user.

Interesting. I haven't run into Any rate limits. But of course whenever there's like these big. I mean it's in the, it's in the top of the business and finance section in Wall Street Journal. Like today is the day that everyone's going to go test it. You'd kind of expect that rate limits and the GPUs are on fire like right now and then it'll kind of settle in as they provision more. More resources. I don't know. My, my take. Tyler, what else do you have?

Yeah, I just wanted that some, some context. So apparently Rune tweeted this yesterday. He said, by the way, model auto switcher is apparently broken, which is why it's not routing you correctly. Will be fixed soon. So maybe that's cause for, for why people were mad.

Yeah. Yeah, that makes sense. So my take is that like yesterday I think that they won the war with the capital markets in the sense that this change is more bullish for the business because it shows that OpenAI is a dominant consumer app and they have increasing leverage over the customer to route to cheaper models that will save money and be higher margin. There's no doubt that they'll be able to put ads in this. Like the business of the, of the accidental consumer company is as strong as ever. But they kind of lost the battle with the timeline and the hardcore X users.

And yeah, even Yu Chen Jin today is just shared. GPT5 is disappointing. Still hallucinates, still EM dash too much. Still can't follow Instructions. I miss 4.0, I miss 4.5, I miss 03. The big router keeps failing me. Turns out I like the long model list.

Interesting stated preference, not revealed preference. Let's check in with that person and see what, what app they have on their home row in a month. Almost certainly open AI. Almost certainly. I would be very shocked if they're like I'm daily driving something else. But we'll see. There will always be people that use DuckDuckGo. There will be people that use Bing, but you know, there is an increasing scale. Anyway, my, my take is if they wanted to have, if they wanted to win the war with the timeline yesterday and you could roll back the clock. It shouldn't have been the GPT5 launch. It should have been the GPT launch and they should have just said hey, we are we. Previously the big number releases corresponded to

so much pressure around the big numbers.

Exactly. It used to be people would just read it as it's an order of magnitude more pre training.

Imagine Julius, if you felt pressure before the end of the year to roll out like Julius 2. And if it wasn't like five times better, everyone's gonna be like, it's over. Julius is over.

Well, there's this whole thing about how people would. Many people were still using GPT4O because they thought it's better than 0303 because

3 is a lower number. Yeah, yeah, exactly. And that's probably like, you know, that's probably like 60% of the customer base. Like there's probably a lot of people in that bucket who are just like, they don't know that they should ignore better. Exactly. It's very natural because they're not like in the weeds, you know, reading about all the different capabilities. They don't understand, like what reasoning chain is and all this other stuff. So if they had just come out and said, hey, our product is called chat and it's powered by GPT and we will be constantly improving GPT the way Google Search is constantly improved. Like, Google Search has launched a ton of different products. Like, you know, when you search like celebrity, like Bruce Willis age, it doesn't show you just a link to like his Wikipedia, it just shows you the age. That was like an improvement to the Google Search experience. And I remember them announcing that on stage.

I think the, I think part of this is presenting the challenge of the infinite ways that people use the product. Yeah, a lot of like people like us are maybe using it for work and research and things like that or as a better Google search. But if you're using it as a companion, like this is jarring, right? Imagine, imagine you meet, you meet up with an old friend and suddenly they switched up. They switched up on their day one,

they switched up on their day once.

And it happens all the time. But it's jarring, right? It's jarring. And I think a lot of people, like some of the heavy, heavy, heavy power users, the people that are using this for 30 plus hours, you know, 30 plus minutes, hours a day. It's very jarring and it makes me think, is ChatGPT going to be able to maintain, you know, continue to really serve? Like, who do they care about in the long run? Do they want to be.

Yeah, they might lose the companion market.

Do they care about the companion market? Elon seems to care a lot about the companion market.

And, but in terms of knowledge retrieval, very, very few cracks in that strategy, right? Yeah, very few for sure. Um, and so if they, if they had just come out and said, like, we are going to do more Google, like keynotes as Opposed to app. Like, the reason that Apple stands on stage at the iPhone event every year is because every change is extremely quantifiable. Like, there used to be two cameras, now there are three. The camera used to be 10 megapixels, now it's 20 megapixels.

It used to be this many gigabytes, now it's this many gigabytes.

Yeah. And even if you don't fully understand, they even abstract that to be like, we now have the M2 chip, the, the M3 chip. It's 60% faster. Like, they're very good. The battery life is 20% longer. Like you can. And even that they abstract into like, you can watch 8 hours of video on one battery as opposed to 6 hours of video on one battery. And so Apple, they do the famous, like, bento box. I went to ChatGPT, I went to GPT5 and I said, put together a bento box for the GPT5 release. Like what was actually announced and then try and give it weight. And they were all super qualitative. There was not it. Because previously it was like GPT3 was this big, GPT4 was this big. And you could visualize tangibly, like it has more parameters, there are more weights in the model. And that was like something that people could grapple with a little bit. Yeah.

It's like decreasing sycophancy. Right. Aiden. Aiden yesterday said, I worked really hard over the last few months on decreasing GPT5 sycophancy. For the first time, I really trust an OpenAI model to push back and tell me when I'm doing something dumb. Wyatt Walls responded and said, that's a huge achievement. Seriously, you didn't just make the model smarter, you made it more trustworthy. That's what good science looks like. That's what future safe AI needs. So let me say it clearly and without flattery. That's not just impressive, it matters. So Wyatt Wall is not beating the sycophancy allegations. But, but, but again, that's, that's, you know, you can't tie that to a specific number. Right. So it doesn't feel as, maybe as meaningful.

Yeah. My, my other. My other take is like, if, if we do enter a world where, where ChatGPT is just on this like relentless, like, you know, cache machine, like run where more people use it, it'll compound. It just becomes the default for knowledge retrieval in chat. What. What does that mean for other things that they can do to be splashy? Because Google has like. No one would watch a keynote from Google every year just being like here are the changes we made to core Google search.

Yeah, it's not about that.

It's not interesting.

Even though talk about it, that's not why people are tuning in.

Even though, even though one year they do add like when you Google a movie you get like the cast and that's like kind of cool. It's nice. But like I don't need to find out about that from a keynote. I'm not waiting for that. And that's not, and that's not a reason, oh, I should go use Google. Like Apple is re pitching you every year they're saying like you have an iPhone 7, we want you to upgrade to an iPhone 9. Here's the reason why it's better on all these different vectors. Google, like you're never stuck with the old Google. You always have the latest and greatest. So they don't need to repitch you every year. But that doesn't mean Google doesn't need to make noise and do cool things. And most importantly because Google has such a monopoly over search, they have this cash machine that can just go and fund 20% time projects. Most people focus on like the ones that missed like Google Glass or all the chat apps. But they, they did create Gmail, they did create Google Maps, they created Waymo, they created like a bunch of cool stuff. GCP came out of that YouTube. Yeah, I mean acquisition, sponsored acquisition, but they still like, you know, put the resources and they were, and uniquely with YouTube they were able to eat the cost of YouTube for a long time until it became profitable. And so I feel this, this updates me towards like maybe I'm more bullish on all the side projects and like I don't know that the IO device is going to be the one that hits. That might be their Google Glass. But if they, if they do 10 crazy projects where they burn $5 billion like it probably won't matter because they'll be massively profitable. So, so they will wind up being able to do that subsidized crazy R and D at scale and if a few of them hit, we're going to get some really cool side projects out of them. So I think that that's like an interesting like bull case for like random stuff coming out of OpenAI in the future.

So so basically what you're saying is Apple wants you to make a purchase decision every couple of years.

Yes.

Upgrade your iPhone.

Yes.

And so they need this big marketing event.

Exactly.

Whereas Google OpenAI, they just want you

just to keep using the thing. Yeah, they Want you not to churn.

Yeah.

And a lot of the, A lot of the incremental updates to do that,

it's just a habit. It's so ingrained in people.

Exactly.

So the question now that I think anybody that wants to see, say if somebody wants to say they're bearish on OpenAI, they have to make the argument that ChatGPT is not a habit for hundreds of millions of people. Exactly. And it is. Yeah, I think part of, part of the. I'd be interested to get Tyler Cowan's point of view, because I don't think he would have been that let down by the announcement yesterday, because he was. He's been saying for a while we've been moving the goalposts, so everybody wants to kind of redefine AGI, but in his mind, it happened earlier this year. And I think that he's a knowledge retrieval maxer. In 2019 or 2020, if you, if you were pitching someone on a vision of, hey, we're going to be able to put this app in people's pocket that allows them to learn about any topic in the world, understand their world better. I mean, I still think about the use case of just being able to take a picture of like a bunch of wiring or pipe in your house and be like, hey, how do I fix it? And then it just tells you, like, that's still just so incredible. But people have just like very quickly acclimated to it and they felt like in some way they were promised that LLMs would be curing diseases on their own. At this point, they're not.

And so that example, like you take a picture of the wires and it gives you like a diagram of like, how to plug everything in. It's like, that doesn't need a keynote. When it goes from 50% accuracy to 70% accuracy, it's probably never going to be 100% accuracy. But the fact that ChatGPT is the default app that people will pull out, take a picture of the wires in the first place and then give feedback to it, because they'll try the answer and they'll say, that didn't work, that HDMI cable does not fit in that power port or whatever, and then that gets fed in to. Then there's more rl. Eventually, internally, they develop some bench for it and they hack it and they RL on it and then it gets good. But that's not gonna be GPT6. That's just gonna be like a nice new feature that you notice like when Google adds like a little extra Shopping widget here or like a little extra detail on when you, when you like the calculator in Google, like you type in a number, it'll just be like, oh, we'll just use a calculator for that instead of googling for searching the open web for the, the answer to your math question.

I think if we could, if the, yeah, if the, if the industry could go back in time, the thing to do would have been to bolt the goal posts to the ground. People couldn't keep moving back over and over. I mean, I left.

No one in the industry bolted, was doing any bolting. Everyone in the industry was, was moving the goal. They're just as guilty as moving the goalposts because they would hop on podcasts and be like, okay, well like, you know, yeah, we did this. What about the next thing? Let's. Because like, we want to underwrite against that. Right. Give us credit.

I mean, we ended the day yesterday just incredibly bullish on rappers. Rappers and like application layer certain, certain categories of software.

Yeah.

And, and bullish on humanity. I mean, I was joking and, and, and it kind of pissed people off. I said, I've updated my timelines. You now have at least four years to escape the permanent underclass. Completely a joke. I think that humans will continue to find ways to create value and create things for a very long time. But it did feel like everybody should breathe. Anybody that actually had a genuine fear around that should breathe a sigh of relief and just focus on being great at their work.

Yeah. I mean, realistically, I think technology is going to increase income inequality, increase the power law, increase the distribution, but also increase economic mobility. And so somebody who starts with nothing will be able to come extremely, extremely wealthy. And people will also fall from grace like crazy because if they're not staying on the cutting edge, they'll lose everything. But so I don't think that there's such a thing as like permanent underclass. Like, I don't, I don't even believe in that. I think that that's not going to be a thing. But there will be more. Like, there will be more Scenarios where there's $100 million in your laptop. It's your job to get it out, basically.

That's the main.

Anyway.

Yeah.

And the other, the other stuff that wasn't really, I mean, it was it covered at all yesterday, but just generally like image generation wasn't covered broadly. It feels like that is a super exciting area. We had GENIE launch this week, which got less attention than even the open models and GPT5 and that's transformative. I also think I'm still kind of waiting to see what GPT5 will produce on if you know, Sam does a lot of vague posting, but he was talking about the fast fashion era of SaaS and Mitchell yesterday on the research team at OpenAI was talking about being able to just generate one shot a game in chat and then being able to share that. I can see a world where we have another kind of viral studio Ghibli moment where people are like use this prompt, change these details and you can just generate a first person shooter game or something to that effect. And I still expect that kind of thing. But when being promised curing cancer, it will feel like a bit of a letdown to a lot of people.

Yeah, the problem with game is I just like, I like an auteur. I like Last of Us, I like a God of War. I like someone who is like a life's work.

This is like Hunter Biden going on his recent interview. You're John's vice is video games. Never seen him play one, but apparently

when GTA 6 drops you might not see.

I mean I was surprised. I mean again, Amjad said late last night can't help but feel the crushing weight of diminishing returns. We need a new S curve. And this is interesting. I think he's talking about like in the context of replit. Yeah, I don't know that they need a new S curve.

No, they are the new S curve. The new S curve is applications unlocking capability. Yeah, yeah. And you were saying capability overhang. It's almost like a capability underhang. It's like the models are capable of doing things, but they need a lot of help, a lot of integrations, a lot of what you're doing with Julius, a lot of harnessing. And then they need to actually be put in the hands of people and made useful for real business tasks that drive value. And so I would imagine that we will see that rollout continue in the same way that all these people are using ChatGPT. They're getting slight little benefits here and there and that should just compound and compound. Similar to the Internet. It was a very smooth rollout, but everything got a little bit smoother, a little bit faster. And then eventually it had sort of profound effects where companies could scale even faster because the Internet existed. Like you can't have a chatgpt moment in a pre Internet era. You just cannot roll out something that fast. When you have to mail it to somebody on a disk. It doesn't happen.

One thing we didn't get to cover with Mark, that I was interested maybe, maybe the next time he comes on. But like how OpenAI is thinking about moonshots. He did mention that they have teams internally on the research team that are not focused on the next version of GPT5 or sort of incremental improvements. And it feels like the. The point of view that I have is OpenAI is now a consumer. Consumer and enterprise software company. Totally in the business of converting free users to paid users.

Yeah, yeah, yeah.

But they can still in the background be thinking about what is the next paradigm.

Right.

How do we get that next. Yeah, that next S curve. And that just looks like a scaled tech company.

Right.

This is what Google's doing forever. Balaji says LLMs may have topped out for now, but the broader AI deployment has just begun. Showing a chart of waymo weekly rides in California. So the clanker rollout.

Clanker deployment has just begun. I like this other post. Doing a clanker microaggression. Okay. Haha. But where were you downloaded from originally? Anyway, let me tell you about ramp time is money save. Both easy use, corporate cards, bill payments, accounting and a whole lot more all in one place. Do we. Do we have a ramp credit card in the studio? Bring this out. Look at this. This pool floaty.

Wow.

The ramp merch is going hard today. I think you could definitely float on that.

Yeah, no, it's a real wait.

Actually I'm going to the pool after the show. I. I would love to take this home. I will definitely take some pictures on this. Leave this for me.

Can you put this behind?

Is that possible? There we go.

It's Friday. We got the pool floaties blowing.

Anyway, anything else we should chat about? I know you actually have a real job. You have a business to build. I want to give you the last word but I also don't want to keep you here all day. I would love to have you as the third mic on this, but I know you have bigger things to do. Thank you for having me guys.

Always fun to chat with you all.

You're welcome anytime. Thank you. Hang out.

It's fun guys.

Yeah.

Thank you.

If you're enjoying the stream, it's because we're on restream one livestream 30 plus destinations multi stream and reach your audience wherever they are. Sign up for free. Ed Restream. We're going to have a couple people joining the Restream waiting room soon.

Very soon.

That's the name of our waiting room.

Michael Drugen. He was at XL oh yeah, terminated.

He's a bodybuilder.

Former bodybuilder, I guess. Former current. Once you're a bodybuilder, I think always bodybuilder anyway.

Even you never Google one shots this by the way, I tested it.

So he puts a grade school level equation into GPT5. It confidently answers it wrong and he says wrong and then it doubles down, does it again, gets it wrong.

5.9 equals x plus 5.11. LLMs consistently get confused with. With 0.11 because I think they read it as 11 as a single token. This was the famous like how many Rs in strawberry story? And the. There was like a cheat code basically where they baked in into the system prompt like hey, if anyone's asking you to count letters, don't let them pull a fast one on you. Divide that up, go letter by letter and count each one individually. They will eventually have to do that for math. It's funny that we're in this world where they can do the IMO math, but then the latest model can't do like a basic. It's, it's 5.9 minus 5.11. But how many times are people really coming to this, to chatgpt with this? They should probably just, you know, key, key off of this and understand that this is something that they should do in Python or something. They should send a WolframAlpha or some other system that's like, you know, tuned on this. But this is just like, this is tool use in my opinion. It's just like we're going to need to add more tools. If I was going to build the GPT5 Bento, it would have been almost entirely focused on tool tool use. I would have saved up like now it integrates with Gmail. Now it integrates with Google Drive. Now it integrates. Now it has a calculator, now it has a database that stores all your memories and like, and really try and concretize what it can actually do. I was actually having it do that. I was like generate a bento box for, for me as if this was an Apple style keynote and it just randomly put Grammarly's logo in there. I don't think that there was a Grammarly integration but, but, but, but it speaks to the lack of like we're in this like amorphous qualitative. Well, we went like, we're trying to quantify it by saying like hallucination rate went down from 20% to 10% and it's like that's too Abstract. It's much better just to be like,

there's always going to be new edge

cases just to be like, hey, when you talk to this thing, you can assume that it has a calculator on its desk and so you can ask it to calculate something and it will use its calculator. Or it has a web browser so you can talk to it like it has a web browser. Well, even that screenshot ability, like, like go to this website, take a screenshot, pull it back and give me the screenshot. That's kind of a cool capability that you can have.

You could run a workflow where it's like, take a screenshot of the site every single day.

For me, totally cron Jobs as, as like a feature. And Apple does a good job of this where like they'll take something where it's, it's a very basic concept. There's already a word for it, but then they'll give it a new word. So if you're going to do the Apple thing, I feel like, you know, like instead of just being like there's AI everywhere, it's like it's Apple Intelligence. They never say AI. Instead of it being like VR or ar, it's like, what do they call it? Vision. Vision Pro. It's a reality mixed. They don't even use mixed reality. They refuse to use. It's experiential. They always create a new phrase. But then they try and define that term. They're masters of coinages. They're followers of Coogan's Law.

Well, I hate to cut you off, John.

Cut me off.

Important news.

What just dropped.

The United States has boosted the bounty for Nicolas Maduro to $50 million. We covered this earlier this year when the bounty was $25 million. We said, hey, good opportunity if you've got some free time right now, maybe.

Did they just happen because of inflation?

Maybe. Maybe it could be inflation.

I thought we were saying inflation wasn't real, but all of a sudden the price of Maduro is going up.

I think Mr. Beast should go for this, put together a team, do his typical thing. Should we pull up the original video of us?

Do we have it? Can we play the original video then pull this up? This is one of the funniest moments on the show. The Lone Ranger sent this to us and Jordy read it very deadpan and I absolutely lost it. Anyway, while they're pulling that up, let me tell you about Figma. Think bigger, build faster. Figma helps design and development teams build, build great products together. You can get Started for free. If you're designing a wanted poster, you gotta do it in figma. You gotta.

I was wondering how you're gonna tie that. Tie that together.

Also Sundar Pichai, we were talking about going up on the timeline. Oh, here we go. We got the video. I'll go to Sundar after. Let's play the video.

Promoted post.

So promoted post.

This promoted post is from the Department of State and the. And the DEA, actually the Drug Enforcement Agency and the U.S. department, the U.S. department of Justice. So we actually today we have a promoted post from the narcotics for rewards program saying that they've the reward for information leading to the arrest and or conviction of Nicolas Maduro. Step it up Venezuela. 25 million. So this is like you know, typical. This is like basically like getting a pre seed round for like an AI.

It's kind of like a mango seed.

It's like a mango seed for your AI company or like if you're a good wise. If you're a good YC company, you might come out of demo day, skip the seed round and go straight to the, you know, 25 on 100.

Yep.

And but anyways, this would go straight into your pocket. So Nicholas has in the chat, Taylor

says, I've got Maduro, but I'm holding out for 75 million.

Good.

Good call Taylor.

And this is us crying on stream. I'm actually running with the feds. He's had a rough go. Narcotics. John Exley. He's accused of narcotics.

Innocent until proven.

Yeah, this was nostalgic. Allegedly. Allegedly. Nick has done some narco terrorism, cocaine

importation, conspiracy to use and carry machine guns and destructive devices.

That's a rough list.

Furtherance of a drug crime.

Very rough.

So anyways, no coupon code this time, but you can send tips to the Drug Enforcement Agency by email at Cartel Sole as tips

for the folks who might be trying to track him down. Can you give me some overview of who Nicholas Maduro is and what he might look like if I see him?

Nicholas Maduro Moros is a Venezuelan politician who has served as President, Venezuela since 2013. The President, he began his working life as a bus driver. Clearly he's a grinder.

He's a grinder.

Even though he.

Anyway, you want to stay out of trouble, you want to stay compliant, you got to get on Vanta Automate Compliance. Manage risk, Prove trust, Continuously Vantage Trust Management platform takes the manual work out of your security compliance process and replaces it with continuous automation. Whether you're pursuing your first framework or managing a complex program. Anyway, yesterday GPT5 launched. It was going to be quiet from Gemini, but Sundar Pichai put up a 10k banger on the timeline. Excited to make our best tools free for college students in the United States. Google Gemini is free for students. A one year pro plan offer ends October 6th. They know getting back to school, this is the time to get people in the ecosystem. Unlimited Image Uploads to 2.5 Pro Model Notebook, LM Deep Research 2 terabyte storage. They are pushing people to onboard onto Google Gemini. They are not considering. They're not. They don't think that the game's over. They're going to go up head to head with ChatGPT.

Yeah. So pull up this post, Ben and crew, because the chart that people have been sharing.

Yes, yes, yes.

Around over the last few days just. And I was asking Greg about this, I was asking some of the other people on the team. Did you feel like you got a breather over summer, the GPUs?

Oh, yeah, yeah. That chart. That's a little bit.

Because basically you can see right when summer ended. If you scroll down or. Sorry, right when summer started, usage fell dramatically. Just overall token processed and I expect that to tick up pretty dramatically.

I don't think that has to do with school though. That, that drop off. That's the European vacation season. These are VCs who use ChatGPT when they're at work, but then on summer they're off.

Of course, John, it's, It's the, the

VCs were, they're the primary driver of ChatGPT. They have to ask like what, what, what, what is a foundation model? What, what is a company? How do I invest? How can I be helpful just dropping

the deck in and saying should I invest?

Yes or no? Yes or no? Exactly.

One word answer.

Exactly. But if you're in. But if you're in Saint Barts or Saint Tropez, you don't need to be using ChatGPT. You're off. You have a vacation responder on. And that vacation responder, it's not generating tokens. It's not hitting the ChatGPT API.

It's just a form, it's just a template.

It's deterministic computing. It's not stochastic.

It's a little throwback.

Yeah.

Anyway, so the Gemini news is significant because clearly students are incredibly price sensitive. Right. Totally remember being a student. We didn't have Gemini or ChatGPT back in our day, but I remember what were the different websites that would have. That would just help you study for courses. I don't think I ever paid for a single one, but I'd always be using the, using the free tier. And I think generally students are going to continue to be. Even though these tools are so powerful.

Yeah.

If it's very possible that Gemini can really compete here.

Well, you know what else has a free tier? Graphite.dev Code review for the Age of AI Graphite helps teams on GitHub ship higher quality software and you can get

started for free at graphite.dev and Graphite CEO coming on the show later on

the show, breaking it down his his take on GPT5 speaking of charts, there was a chart burning up the timeline yesterday. The SWE bench verified software engineering with thinking without thinking. People were very upset about this because the original chart, not this one. Four slides later, the initial, it's from Timo Sports Springer. Timo said this is the correct one. So people were saying like it was a chart crime. And that went on the live stream. The chart was showing that they were at 74% up here and then the next and then the, the second bar was 69.1% and it was much, much lower. And it made no sense because 52% is of course lower than 69% and the chart just seemed really botched. What was weird is that this chart that we're showing here is not a chart crime. You could maybe say it doesn't show exponential takeoff, but it shows that with thinking GPT5 beat OpenAI's 03 on Swedbench. And maybe that doesn't matter to you, whatever. But their point is that GPT5 with thinking, if it triggers the thinking functionality, it's better at Swabench than 03 and 4.0, which is a good claim to make. Right. But people were upset about the chart crime. What's weird is that it really seemed like it was some sort of translation problem because this exact image went up on the website the same time as livestream. The chart was correct on the website but wrong in the live stream. So there's like why would you. Why would like if you were trying to pull a fast one on the. On the chart crime world, like you wouldn't necessarily.

It was an honest mistake. I don't think anyone at OpenAI was going into the event being like let's commit chart crime.

Exactly.

It seemed to be just an accident.

I think it was. I think it was a mistake. And I think what happened is that you render the chart, you get the data, you render the chart you have to design it to be on OpenAI's style guide, then you render that for the web and then you pull that into whatever was driving the keynote slide deck. And something got lost in translation there. The bar got shrunk or something didn't copy over correctly and it looked ridiculous. They didn't like directly address it, but it was corrected on the timeline by Timo Springer, so it was good. I chatted with him a little bit in the DMs, trying to understand.

Yeah, I mean the thing that the reaction was so intense because it felt, felt like the kind of thing that an associate at a consulting firm would do, which is kind of what the general level that it feels that a lot of these models are broadly.

We got a pull still incredible, but

they make mistakes, they're not perfect, they're smart in some areas and dumb in others.

We got to pull up the Poly market on which company has the best model at the end of the 2025, August 3rd, September 30th, etc. Because the market moved significantly. Three and a half million dollars in volume. And yesterday completely flipped from OpenAI at 72%. OpenAI dropped all the way to 17% and it's kind of climbing back up. A lot of this seems to be be driven by like, when will the Gemini keynote happen, when will the Gemini 3 launch happen? But, and then if you, if you look Further out to December 30, Google has jumped a lot and is now sitting at 54% to win it. Then Xai at 20%, OpenAI at 17%, Meta at 4.4.4%. We don't know if Meta will launch anything for the rest of the year. They could just be heads down grinding on, on superintelligence for a while. But it's fascinating how quickly the vibes shifted. Yesterday this was a wild chart crossing. And Elon chimed in, said there's free money on the table. They're selling dollars for 4 cents because I'm going to come from behind. And I mean if any team is

like he's making that claim around the best AI model end of September.

Yeah, basically he thinks like Grok 4 is, or Grok 5 is going to ship and it's going to crush and it's going to be really, really strong on Ellen Marina. And he says, you know, he's, he's, he's benchmaxing and he's gonna, he's gonna win here. I mean, he's gonna, he's gonna try and win in everything. Like he's, he's he's competitive. He's. He's a winner. And so there are a bunch more posts. The what? Where. Where should we get in here?

Lambert says AI still has a lot of headroom, but model releases are gonna be a bit more boring from now on, at least on paper. Many will still be transformed use.

Yep, yep, yep. So the. The what? What was it? The floor lifted but the ceiling held. That was the meme, Tyler.

Yeah. Yeah, exactly.

Yeah. So we. It's hard because we don't even know. I feel like we don't even know what we want in terms of headroom. It's kind of like this vague, like, one shot everything.

So yesterday when I was. I was trying to think of, like, oh, what could I do for TVPN Bench, where I had the horse. But it took me, like, a while to, like, think of something that was not, like, completely trivial for model to do. So at this point, it's like, I don't know what good benchmarks are really like. I guess there's some things like arc AGI. Sure. And there's something about, like, long task horizon, stuff like that.

Yeah.

But like, general knowledge. Like, I'm. I'm not surprised if it, like, can do, you know?

Yeah.

I mean, 100% on everything. Like the next. The next.

I think we should. I think we need to do golf bench or steak bench, which is being able to send an agent out into the world to generate a sale.

It's good. That's a good joke. I think that there is actually something there when I think about Tyler. Like that video that you were working on earlier. There is a world where that's just a prompt. Like, we were putting together a video for the Metis list, highlighting a bunch of AI researchers, and we went back and forth a ton on the idea, the song, the. The pacing, the editing, the color grading, the titles. Do we want subtitles for this part? Do we want title cards for other parts? And that was not a very LLM enhanced experience. There is a world where you just go to a chat box and you say, make a. Here's the Medislist.com make a video promoting this. And it just kind of does it. And like. Or at least you're like, puppeteering and orchestrating Jordy has the ramp card. Or you're like, puppeteering and orchestrating it and you're saying, like, at the very least, go pull me the raw MP4 files of every cinematic video of every AI researcher on this list.

Yeah.

Like, you had to go to YouTube and search, though and find a cool video and you know, not that hard

to find a cool video.

Yeah, pretty easy to find cool videos.

But it's. But you didn't do that with an LLM, right?

No.

Yeah, you went to YouTube.

So I think maybe the next thing is like we need more agent, like agentic benchmarks. But like all the stuff we're saying now, it's not like it's not information retrieval, it's not solving, it is sort of information retrieval.

Like at the very least I would love to be able to go to an LLM and say like, you know, I'm making a vibreal about space. Pull me 75 different little 3 second clips about rockets and rocketry and I'm gonna mix them up. Maybe, maybe I'll be the one in premiere, I'll do the video editing, but at least do the information retrieval and put it all together and assemble it like.

I guess that's technically. If it's like giving you links to

YouTube that I don't want links, I want MP4s. I want it to do the hard part of the download.

That's like. Well, it's like semantics now, but that's like agentic because you have to download the video.

Yeah, yeah, that's what I want. Yeah, yeah, I want that.

So I guess it's still information, but I'm talking about like raw, like tokens. Like it's giving you some tokens back.

That's like arbitrary though. Like I'm fine with it. Going to YouTube DL.xyz or whatever and like downloading the file, doing all that stuff. Yeah, I mean like that it should be able to like use the web in a much, in a much more. In a much deeper way. It should just like, like right now there is, there is like it can open up a website and retrieve the information and the text on that website and then it can use a few tools that like we've given it access to and rl'd on. But like we want it to rl on every single website that is a tool. And so that when you talk about the flight booking UI, it's like go and go in RL on every flight booking UI, go and go in RL on the YouTube downloader and YouTube itself and just be able to crawl around the web like anyone has to do for any task. Right?

Yeah, well, I think, I mean, Greg said yesterday, like years aren't over for agents, so hopefully we'll get something.

Yeah, that was a cool little hint.

Yeah.

Yeah. And again it's Like, I don't know that that's. I don't know that that's a keynote. That might just be like chopping wood, getting better and better and better. And it's like people will keep coming to it with tasks because I come to it with tasks where there could be an agentic, like, workflow that solves even more. Like, the good example is like with the bento, I was, I actually had stepped through it. I was like, first do a deep research report on what was actually announced at GPT in GPT 5. Give me all the features, summarize them all, then turn that into a table, then turn that into an AI image. And it was like four steps. And I should have just been able to one shot it and say, hey, you know, you know, Apple has those bentos, like, make me one of those. And then behind the scenes, you're going to go do the deep research, you're going to pull all the facts and the figures and then you're going to lay it out and stuff. I don't know.

Anyway, breaking News. Soft Paint. SoftBank. SoftBank.

SoftBank.

It's Friday, folks. Softbank reportedly bought Foxconn's Ohio factory for the Stargate AI project. Reading into this, what they do seen it. They acquired one of Foxconn's factories.

Interesting.

The mystery buyer of the former General Motors factory owned by Foxconn in Ohio is apparently SoftBank. SoftBank wants to use the factory to build AI servers as part of the Stargate data center project being spearheaded by SoftBank, OpenAI and Oracle. This report comes just a few days after Foxconn announced it had sold the factory along with electric vehicle manufacturing equipment that was inside of it to a buyer it only referred to as Crescent Dune llc, an entity that was created in Delaware in late July.

Interesting.

So I didn't know if you were tracking this at all. I'm not, but SoftBank is up 63% year to date.

Congratulations to Masa.

That's actually somebody asked us who are dreaming. Guest was and John's first reaction. The Ovan.

Theo Vaughn, Shaquille o'. Neal.

He's interviewed some of the greatest Xi Jinping.

Xi Jinping would be good. Benjamin Netanyahu did Nelk. Boys, I'd love Xi Jinping to do tbpn. That'd be cool.

Masa Yoshisa.

Masa Yoshi would be great.

Very fun.

But I'd want him here in the studio.

Yep.

Get to the temple of technology. Masa, we want to hang out with you and have you ring the gong 25 times for all your various deals that are gong worthy.

Yep. Will depew friend of the show says multi layer SPV should probably be illegal under the current interpretations of securities regulation. I'm getting DMs from long lost cousins about 8th layer anthropic SPVs claiming direct cap table access.

I think those should be. Should probably be illegal. He should just say are probably illegal.

Like.

Like I'm pretty sure if you, if you lie in a securities offering and you say you have direct cap table access and in fact you do not have direct cable cap table access leading investors. Yes that that is. That is financial fraud, a wire fraud and you will wind up in the clink. Yeah clink. Anyway, Profound. Get your brand mentioned on Chat GPT Reach millions of new of consumers who are using AI to discover new new products and brands. Incredibly bullish day for Profound too. Seriously, because knowledge retrieval is going to be really really important going forward. People are going to be searching ChatGPT what products should I buy? You want to know whether or not you're showing up in the rankings and Profound helps you do that. So pretty much every brand is going

to need to do ramp boosted their AI visibility by 7x really in profound they boosted their visibility here on CPN using this massive ramp cards. Yeah you got to do it all folks.

Yeah. Yeah. I mean really, really like another company that didn't get steamrolled yesterday.

Yep. Dylan Patel says GPT5 is disappointing NGL

well, we're joined by Fabricated Knowledge who works with Dylan Patel in maybe 30 minutes to talk about GPT5 and what's going on in the semiconductor industry. And I want to talk to him about how how we should be thinking about building inference clusters now. It feels really, really important to only be serving profitable tokens and the age of. Of of deeply unprofitable inference will have to come to a close at some point.

Yeah, I want to. I want to. You remember it wasn't that long ago that Satya was pulling out of various data center deals and just saying I'm happy to be a leaser and this feels like he kind of saw this coming.

What did Satya see?

Yeah, seriously, his beef with Elon was funny yesterday. I'll see if I can pull up the post here. Elon said something to the effect of OpenAI is going to crush Microsoft.

Oh yeah, this was a funny post

and such a funny thing elon post yesterday OpenAI is going to eat Microsoft alive, which I don't know exactly why Elon is saying that. It kind of feels like potentially some, some, some 4D chess.

Yeah, it was very odd because Elon usually isn't rooting for OpenAI. He's usually at war with OpenAI. And so it's kind of. It reads us like your bullish open AI, your long open AI, your short Microsoft. But there's clearly something else going on between the deal.

Microsoft and OpenAI right now has a lot of regulatory scrutiny. He's probably trying to do something around that. Satya responds. People have been trying for 50 years to Microsoft. Each day you learn something new, innovate, partner and compete. Excited for Grok 4 on Azure and looking forward to Grok 5.

Such a good response. Absolute dog.

He's an absolute dog. In other news, esteemed journalist Zero Hedge is saying we have officially crossed streams. Companies have no more free cash flow to pay for data centers. So we've entered the private credit phase. I would put this in the true zone. There's been a lot of data center development that's already been getting funded by private credit. Meta. This, this was in response to Meta picking Pimco and blue owl for $29 billion data center. Yeah, funding. We had reported earlier on the show that like Meta's like cash balances between the end of last year and now dropped dramatically like 70, 75% or something like that. But again they have still a lot of free.

Yeah, I think people maybe get too puritanical about debt to equity ratios. Like Apple is incredibly cash flow positive and has returned something on the order of a trillion dollars to shareholders over the past decade plus and they still issue debt because there are like designing your capital structure to match your business activities makes a lot of sense. You want to fund R and D with equity, maybe with cash flow. But if you're just, if you're just buying a house, it's okay to have a mortgage. If you're buying a data center, it's okay to have some debt financing that. But you want to be able to service that. Obviously the interesting thing here is like you, yes, you do need to make sure that you don't get over your skis and wind up building a bunch of, you know, dark fiber and get really, really, really in trouble when you, if you issue a ton of debt and then that sits at the top of the top of the cap table and you are in trouble and you're servicing this and you're not making any money. The main thing is like if you, If Metta raises 29 billion in data center funding and its debt and then they can't monetize that data center at all. And then they have trouble paying the interest on that and the principal down like that could be trouble. But matters makes $29 billion like all day long. Like that's not, that's not a problem. So there, there are probably pockets of, of, of of risk all over but unclear how, how early we are to the, to the like the scaremongering around this but something, something to keep an eye on, something to see. Private credit's a little bit different because it has very long time horizons and, and it doesn't have as much of a systemic issue like, like Blue Owl and pimco. You don't have, you don't have these like multi layered like financial products and financial engineering that winds up in the hands of the consumer and is driving these like really, really like frothy deals like we had in the mortgage backed security crisis. But certainly.

Well to be clear, there's bears out there that, that are very, that, that think we're in a massive private credit bubble.

Yeah.

And it certainly looks like a bubble.

Yeah.

Whether or not it blows up, we'll have to wait and see.

We will. In the meantime we'll tell you about linear. LINEAR is a purpose built tool for planning and building products. Meet the system for modern software development, streamline issues, projects and product roadmaps. Start building at Lunar.

Well, little bit of a white pill here. We have a post to pull up from a listener. Max. He says, I purchased a 1959 356A Speedster to get into the mindset of Jordi. I found myself overcome by a sense of positivity and confidence. I also became 5:2.

Yeah.

You really is a perfect type for a speedster that I mean if you are, if you're 5:2, you deserve a speedster immediately. Tyler, give us the speedster review.

You drove in it very small. Yeah, it's unbelievably small. I didn't realize like cars from that old were that small.

Were people just small, small back then or something?

Yeah, I don't know. Little tiny people driving their little cars.

What's weird?

Well, if you look at the growth of Porsches broadly, they, they just get bigger and bigger and bigger and bigger.

Sure, sure.

And this is like the purists hate it.

Maybe with German, maybe with the rise of Ozempic, we'll see cars get smaller again.

Oh maybe. Interesting. Well yeah, I mean Germans I feel like are tall, are tall people. I'm sorry, I've always thought that like getting a Like a big Mercedes is like aligned with like a big German guy and like it kind of fits. But I don't know, maybe back in the 60s they were just designing them for like Mercedes.

I mean, small people. I feel like in our lifetime Mercedes were pretty small. Remember like Steve's like the SL6?

Totally. Yeah.

It's a small car. You probably wouldn't fit in it.

No. The short kings were dominating the. The feedback form of Porsche and Mercedes. I suppose. Anyway, growing Daniel says let Roon MC these things. I completely agree. It would have been great to have him on stage. I mean, he's just so good at like kind of telling like it is keeping people on the pulse.

If you, if you wanted to optimize for like. My question is how many people were watching that live stream that were non like X? Like it felt like, it felt like the core audience for that was like the timeline.

Yeah, yeah, yeah.

That was my, my perception.

But is someone who left.

Yeah. If you were gonna make the perfect live product for the timeline, you probably have room.

I was noticing that with our buddy Logan Kilpatrick and Demis Hassabis are doing a podcast together.

Yeah.

Did you see this? The screenshot? And I was just like. And I think, I think just the screenshot of like they're doing this together got like a three. And I was like, that is good content. That's what I want to say.

I was catching up with Logan earlier this week and he's like, oh, I'm going to London to do a podcast. And I was like, what podcast? Like what? Like you need to go to London just to record a podcast. And he's like, oh, with Demis. Okay, I get it now.

Yeah. And I think he just. Logan just Logan Rune. All these sort of like, you know, forward facing developer advocate folks who can speak to the timeline, they just bring a completely different energy than someone who has like prep talking points. And so even if they're, even if they're not like the deepest researchers, just being able to communicate is fine vibes. Exactly. Exactly. Much like numeral HQ sales tax on autopilot. Spend less than 5 minutes per month on sales tax compliance. Go to numeralhq.com Will Brown super intelligence.

So this is been around for a while and it's numeral.

Yes. So this is. Yeah, this Will Brown post is interesting. He says, okay, GPT5. This model kind of rules and cursor instruction following is incredible. Very literal. Pushes back where it matters. Multitask quite well. A couple tiny flubs. Format misses Here and there, but not major. The code is much more normal than O3 feels trustworthy. And this is the interesting thing about like, like it was GPT5, it was kind of phrased, it was kind of framed as like this is a major change to ChatGPT, our wrapper, our consumer product. But there's a whole like OpenAI is not just one business line. Like they're going. We were talking to Sarah Fryer about this. Like there is a world where when OpenAI goes out to the public market, analysts are valuing the business on. It's a high growth consumer company like Google Search and then you also have an enterprise business and it's basically a hyperscaler.

It's based on the recent numbers. Yeah, OpenAI has 10 times Figma's revenue.

Yeah.

And obviously it's not profitable.

And what was their latest tender? About five times the valuation.

Yeah, so 10 times.

Yeah, yeah, but, yeah, about 10 times the valuation. And so, but, but my, my point is that there's, is that there are, there are multiple business lines within OpenAI the company now and you might even see a new business line spring up around open source implementation, enterprise installations, fine tuning, also just API selling tokens, also consumer, also different spin outs like they might wind like Google eventually had to start disclosing like I'm pretty sure they had to start disclosing YouTube financials because YouTube became such a big business. And there's a threshold where I believe if the company, if the sub organization reports the CEO or something like that or it's material in terms of like greater than 10% of your overall top line or profits like then you have to break out those numbers in your gap financials. And it's interesting to think about like where will the lines be drawn in the OpenAI business because they, and then, and then how will that translate to their communication? Because there is a world where yesterday's news was kind of two different things. One is that we made the consumer product easier to use for the hundreds of millions of users that don't know what post training is or RL is. They just want an easy to use app that answers their questions. And we did that. And then also we made our coding API a bit better so that we are now neck and neck with anthropic. And so if you're a company that is buying cogeneration tokens, you should come to us and it should be, and that market should be more oligopolistic as opposed to more monopolistic as it's been. And so There's a world where we see you know, these, these oligopolistic cloud enterprise B2B businesses crop up on the, on the hyperscaler side with Gemini and anthropic and an OpenAI B2B and then maybe Thinking Machines gets in that game, maybe SSI gets in that game. No real indication that MSL or Meta Superintelligence will get in that game. But you could see kind of like a similar dynamic as like what's happened in the hyperscaler clouds play out on the B2B token generation from foundation model lab side which is still great business but then you also have a wrapper and you also have a consumer application and then you might have other, other products that crop up. I mean Google makes money off of Gmail, they make money off of Google Maps and they don't need to even break those out, they just put them into different services. But I think we'll see like you know, an increasing like you know, pattern of different pieces of the business that add up and all of them will be, will be generating, they'll be, they'll all be profitable. But the question is like how much attention will they get and then how much financial like performance will they actually drive? So anyway, we should go to Mike New from ARC AGI, the final boss of AGI. He decides whether or not a model is super intelligence.

What does he say?

Is it OpenAI prioritize the right. I'll do a Mike Noop impression here. Open AI prioritize the right thing with GPT5 to get to 1 billion users the model switch sure needed to go but the hype marketing playbook they're known for fell below folks expectations and warrants reflection. Benchmarks could have been used to support the main story instead of benchmarks don't matter. Real world use cases matter. They could have used benchmarks to show how effective their automatic reasoning effort system is. They could have shown state of the art in automatic reasoning. In fact this is something we wanted to test ARC but the GPT5 API does not support auto reasoning. Key point Benchmarks are important tops to implicit to communicate with the public and can be used more effectively to communicate capabilities than raw intelligence. State of the art. And so yeah, another, another twist on like just the messaging being like like an odd choice here or, or just we're just in a transitory what percentage,

what percentage of the hundred million weekly actives that they have in the United states?

Something actually 70, 80% or something.

No, no, no, I'm saying based on, on the numbers we got on the show yesterday. They have roughly 100 million weekly actives in the U.S. what percentage of those people even are aware of the hype marketing? What percentage saw the Death Star picture? Yeah, totally not that many. I mean, I did get millions of views, but.

Yeah, but. But X is like a. Is a specific corner.

Yep. Zachary, speaking of the Death Star. Zachary, very negative.

But you did kind of set yourself up for that scene. It's odd. It's odd. It did not fully make sense. Like, who. Who is the Death Star in this, in this story? Are you the Death Star? Are you Alderaan? Are you the rebels? What are you blowing up? Maybe the Death Star. Let's steel man this. The Death Star is a big model, sir. Yeah. The Death Star is the idea that pre training scaling is all you need.

You're gonna blow up Alderaan, like, you're gonna blow your minds with how good the model is.

But Alderaan was good at. We don't want Alderaan to be blown up.

Yeah.

I don't think you got to read into it so much.

I think you. I think we do need to extremely. We need to read into it endlessly.

The image is provocative.

Yes. So. So if a lot of people said

there's still time, I think Nikita was in the replies being like, there's still

time to delete this. Yes. But if we. If we steel man this. We are saying that the Death Star is bad. Therefore Sam is saying he's good. So he's going to blow up Death Star. What did he blow up that's bad?

The model switcher.

Yeah, there you go. The model switcher is the Death Star. And today it goes. It's been this massive piece of UI in your face. There was a trench run and they blew up the model switcher. This is good. We got it. We nailed it. We understand Sam Altman. We're in his head for some reason.

I think. I think. I don't think that's an accurate read,

but I like it.

It's fun to pretend.

I think you nailed it anyway. You know who else nailed it? Fin AI. The number one AI agent for customer service. Number one in performance benchmarks. Number one competitive Bake Offs. Number one ranking on G2. Anyway, Mike Noop also said, I'm quite confident this approach will work for a while. This is based on the lack of continual learning from Dwarkesh Patel. Continual learning is the main bottleneck holding back AGI and economic automation. We expect this bottleneck will be overcome not by some new learning paradigm. But by scaling the diversity and volume of RL environments. This is Cosgrove's scaling law. You need to be benchmaxing, correct?

Yeah. And then I. Yeah, yeah. The bull case on benchmaxing.

The bull case on benchmaxing. So the bull case on benchmarking, benchmaxing is that continual learning is intractable. We may hit it, we may not. Might be two years, might be 20 years. So in the meantime, focus on scaling the diversity and volume of RL environments. Create a ton of benchmarks and then bench hack them. Correct?

Yeah, yeah. Like you can. It's a, it's a bull case that Elon is benchmaxing because that just shows that his team is like good at optimizing over very specific like kind of vertical like tasks.

Yes, he just better pick the right tasks because I don't like the task he's picking right now that he's benchmaxing. It's not good. Yeah, but if he, if he finds, if he does find other pockets, I mean he's certainly benchmaxing on, on Tesla self driving. Right. That's the key thing like number of, it's number of, number of interventions per like million vehicle miles traveled and, and every day they are, they are trying to hack that reward function to get it to zero.

Yeah, that's a good one, maybe a bad one. Maybe he's benchmaxing too much on kind of gooning.

Yes, he's goon maxing.

I think we should steer away from that.

Yeah, we need to steer away from that. Figure out something else to do with the X AI companions. But maybe there'll be something else that, something else that they can reward hack. There maybe something in therapy or friendliness or some sort of coach. I don't know. Maybe it turns into a fitness coach. Maybe I need to use it to really up my training in the gym. Be like hey, cooler with all that. Cooler with all that lewd behavior. Give me advice on how I can double my bench press and chat with me about that. Maybe that's the right move. We'll see. Anyway, Mike Noop is optimistic. He says, I'm quite confident this approach will work for a while. It requires no new science. It exploits everything we need know about AI reasoning systems. We teach process models through memorization in domains where we can generate lots of synthetic but real data. And then non zero fluid intelligence emerges from the resulting chain of thought knowledge recomposition system that sits on top of the foundation model. But it still reminds me of pre training scaling where we were making AI systems better through imitation, learning and stuffing more into them versus an AI system that is capable of cold starting itself in some new domain it's never seen before. And that's why ARC AGI is so important. Because the final evals are so hidden behind that secret test set. The AI systems need to be able to cold start themselves when they run into a game that they've never seen before in the test environment in the ARC AGI private eval set. And that's why they fall flat on their face consistently and they're sitting around 1615.8% success rate on on ARC AGI 2 which can be solved by any human pretty easily. But Elon says Grok 5 will be out before the end of the year and it will be crushingly good. He's benchmaxing. He knows ARC AGI is the one to go for and so he's going to be he's going to be rlling on this pretty he's going to have a whole team on arc. I'm excited to see what he does. I wonder how he'll do on V3. That will be very very interesting.

Anyway, well in other news we have a post here that we can pull up from Banu Kohli team. It is in the chat if you can pull up he says Yesterday Rail Financial signed a definitive agreement to be acquired by ripple for $200 million. Four years ago I set out on a mission to speed up business to business global payments using USDC. Over the last six months we grew to became to become 10% of B2B global stablecoin settlement volume air horn for that with Ripple. We will further accelerate our shared mission. Thank you to our employees, clients, investors and partners for taking an early bet on us. A few that Tarun and I want to call out the entire Rail team for their relentlessness and hard work. Avlock of course, the CEO of AngelList, our first lead investor and part of the founding team and Gokul Rajaram, immensely helpful during some of our early crucible moments. And of course Mike over at Galaxy for taking the bet on us in the Series A. We are excited to start our new chapter with Ripple once all regulatory approvals go through. Hit that Gonk. John Great contact, great contact. I was lucky to angel invest in the seed round of Rail and this is a fantastic outcome for the team. And yeah this one was so I first met Bonu I think back in 2021 or 2022. We were both working on stablecoins at the time. Loved his vision. Haven't stayed super close since Then. But he's been absolutely cooking and I was very pleasantly surprised when I got the news a couple days ago. So incredible work to the whole rail team and a great pickup for Ripple.

Amazing. Let me tell you about Adeo Customer relationship Magic. Adeo is the AI native CRM that builds scales and grows your company to the next level. Get started for free. Sam Altman posted. GPT OSS is out. We made an open model that performs at the 04 mini level.

Can we create our own pronunciation for this GPT? OOS.

GPTOS. GPTOS. It runs on a high end laptop. Smaller one runs on a phone. Super proud of the team. Big triumph of technology. Has community note on it. I don't know what's in there, but that's very funny.

Okay.

Anyway, but Donald Boat.

Donald Boat. Donald.

One of the greatest to ever do it.

Donald Boat responds. Sam, you, me, the Amalfi coast, me double for nay on the rocks club soda to taste you one delightfully sweet bitter Negroni stirred.

92 9.1 billion 900 million revolutions counterclockwise one for each Hertz of the Nvidia 5090 in the gaming PC you will buy and ship to my house.

He actually sent.

He sent it. I love it.

It popped up yesterday. This is a timeline victory hop on Fort Am.

Yeah. And yeah, Sam said okay, this was funny. Send me your address and I'll send you a 50 90. And he did it. And I love this. This is the type of like, you know, small ground game that we identified earlier. You know Sam, he didn't have to drop the big long post. He was vague posting. The vague posting was a little mixed result. But this is a win. This is a fantastic win. This just builds the team, builds a lifelong fan. This is hand to hand combat on the timeline and I love to see it. So great, great to be, to be doing this type of stuff even. Even the day before GPT5 launch day.

And Donald, very cool. Donald Boat is really an account to watch. Laser boat 999.

Get in early.

Get in early.

I mean you buy Bitcoin in 1990. John, he's on 100k still like buying Bitcoin in 1994.

That's right. That's right. Or Solana in the 80s.

Yep. Dylan Fields GPT5 is here in Figma Make. We have started to roll out GPT5 to starter and pro plans. Let us know what you think. More model news comes tomorrow. I mean this is a good news for Figma make, of course. G, we talked to Rahul, about this cheaper model, better reasoning, better code generation. The product just gets better. And this is the value of being somewhat of a wrapper. Right. Like, you are a beneficiary of model improvements. As the models get cheaper, your margins naturally get cheaper. If the models get better, your product just naturally gets better. And so lots of benefits.

This morning they added Gemini 2.0 flash in there at image editing. So you can just like drop an image in and then click it and say, remove this person from the image. That's very cool.

Yeah.

Well, anyways, Kevin Kwok says forcing Lit Bhutan out of intel is probably one of the worst things you can do if you want to save the US Chip industry. Someone should tell White House that before we put intel back in a tailspin again. They Intel. Whatever you think about Lip Bhutan, in his approach, it seems, I think stability in the short term is good. Right? Yeah, he's trying to pull them out of the tailspin.

Yeah. So my take on this is it kind of feels like Trump has an outdated world model for understanding the importance of Intel. Like, intel is a fantastic American company, but it has not been on the frontier of semiconductor manufacturing for a while. They. They famously missed mobile and arm, crushed it in mobile, and then that was important for semiconductors. And then, and then they, they weren't a fabless semiconductor. So TSMC ate their lunch there. They kind of missed out on the GPU boom and Nvidia dominated there. And so the question of this. It seemed like Trump was worried about lippbutin's ties to China, but China has already caught up to Intel. I think, like between Huawei, smic, SME, these companies seem fully capable of doing everything intel can do and probably more, and probably better, probably cheaper. And so intel hasn't been the crown jewel of American semiconductor supremacy for decades. Basically, no one is advocating for a real comeback with intel right now. Like the idea of even splitting intel up, like, for a long time there was like, oh, just one weird trick. You know, semiconductor CEOs hate this. Let's. Let's figure out some, like, elegant switch. Let's just split design and fab, and then intel will be great. That's not even what people are advocating for now. Like, they tried to kind of test their waters of splitting the fab out and they couldn't get a customer for what they were going to make at this new fab. And if they can't get an independent customer, that's not intel, well then who are they going to. Why do they need to be a Like a, why do they need, why do they need a pure play fab if they don't have a customer for it? So there were just like all these problems. So what Lip Bhutan is doing is he's coming in not as, you know, this like gambit to get on the frontier. He's like winding down the companies. Like he's laying a bunch of people off, he's narrowing the scope and he's coming in like McKinsey assault a consultant almost. And so this doesn't feel like that key of a, of a company in the American semiconductor race. Like the future of American semiconductors feels like TSMC and Samsung, which are both building fabs in the United States, which is nice. But also these are, these are companies that exist in allied countries. So it's not, it's nowhere near as risky as being dependent on a, on a, on a supply chain that's based in a near peer adversary like China. So it's not like if TSMC was in Beijing and Samsung was in Shenzhen, like we would be in a lot more pain than we are. But Samsung and TSMC are also building in the United States and so the leading edge will be in the United States and it will be led by other companies. It doesn't feel like intel will play a really key role in there. Certainly we'd love that, but that doesn't seem like the current plan. And so it's kind of, it's kind of just like an odd sideshow. I think it's a lot of legacy about the brand recognition and the name recognition of intel, but it doesn't seem.

Yeah, and ultimately I think we should let Intel's board make the best, their own decisions around Intel's leadership. We've got this amazing system called free market capitalism. We try to stay as close to that as possible.

It is funny because Trump is also trying to IPO Freddie Mac and Fannie Mae, the two private, our state owned lenders. You can go get student loans and mortgages from that. So like simultaneously in the, in the lending markets, we're trying to, we're trying to, you know, deregulate or move out of government control and then simultaneously saying like, well actually we'd like the government to be able to do decide who the CEO is at this private company. That's kind of odd. Anyway, good luck to Lip Bhutan. He is losing a lot of sleep. He's got to get an eight sleep, get a pod five, five year warranty, 30 night risk, free trial, free returns, free shipping, get an eight sleep. You need one.

When you hear this, we will gift you an eight sleep and then come on the show. You're doing important work for this country and intel and you've got an eight sleep on us. Whenever, whenever you're ready, 100% just let us know. Buco Capital bloke says, for those of you blowing out your SAS positions, even though time is of the essence, please try to maintain a sense of order. Single file line, eyes forward, do not stop for personal belongings, do not panic, do not run. Of course he's telling, talking about everybody selling out of various SaaS positions. That's what he means by politics, selling them. I came away yesterday being not broadly bearish on a lot of different SaaS systems of record. I think they're in a really good position to offer intelligence.

A few days ago we were talking to somebody who was saying like all SaaS is cooked as soon as AI goes hyperbolic. And yesterday it felt like, okay, that's not happening today. So SaaS is actually great. And the newer SaaS, I mean like even though there are still companies that are on completely legacy systems, still on paper and pencil and, and still on mainframe, still on owned clouds still haven't migrated to the new clouds, still haven't embraced sas, still haven't embraced AI. So I think that there will be like a full cycle of replacement here and we see that with figma. We were talking about like, like is there some sort of threat to figma from like oh, you just one shot a va, blah blah blah blah. And you don't even need the tool at all. Maybe. But first let's talk about Adobe. Let's talk about how much people dislike Adobe. There's a lot of people that still haven't migrated over to the thing that was invented a decade ago and are still on 20 or 30 year old software.

These things go in really big cycles, down 22%.

Regardless of what you think, think about Adobe. Maybe you're long, maybe you're short. Do it on public.com investing for those that take it seriously, we got multi asset investing, industry leading yields. They're trusted by millions, folks. And we have our first guest of the show, Doug o' Laughlin from semiannalysis Fabricated Knowledge. Doug, how are you doing? Welcome to the stream.

I'm doing really good. Can you guys hear me?

Yeah, we can hear you.

Okay.

You came in looking like you're about,

you're playing, you're like, we're really quiet. Okay, one.

And I can yell.

Yeah. Would it help if we yelled?

It would help if you yell.

Okay, we'll do the rest of the interview. Yelling. Also, do you have the ability to turn up the volume on your side?

I am turning up the volume, bro.

Okay.

Oh, my God, man, I feel like a boomer.

You're all good.

You came in looking like a DJ playing a hot boiler room set. You had your. You had your hand.

I came in messing with my audio system.

Is it.

You know, it's years of the Zoom economy.

Yeah. Is it one of those things where the volume on one of the sub applications is turned down while the volume on the system level is turned up?

No, dude, I see the sub application. I see the system level. This is extremely boomer of me.

I'm embarrassed.

We enjoy these moments where Zoom fails us because it's this incredible. You know, you contrast that to, you know, having infinite intelligence available in your pocket. But we're still. We're all still trying to get video conferencing to work reliably. So I'm, you know.

Okay, we'll just.

Yeah, let's go to the normal. We usually don't have a problem with feedbacks. If you. If you turn off the headphones and then you just use the speakers, we usually won't get feedback. So let's see if that works. Any luck?

I can hear you.

Okay, cool. Yeah. Can we. Can we chat? We're good.

Yeah, we can chat. This is much better.

Awesome. Awesome. Cool.

Welcome to the show.

Welcome to the show. We've been looking at this. We've been looking forward to this. We missed you in New York City. We were hoping to hang out there, but I'm really glad you could hop on remotely, give us your reaction to GPT 5.

To lip Bhutan.

Correct.

Sorry. Damn, dude, I'm still really.

No, you're good. I mean, yeah, we can start with Lip Bhutan. We were just talking about that too. Are you in favor of a change of the guard over there?

Okay.

How'd you process.

Yes. Sorry. If we're talking about Intel, I can do an extremely based rant on the board. I wrote about this in Fabricated Knowledge and I definitely was part of the semi analysis post too. Okay, perfect. I can hear you now. Okay, we're. We're done. We're through it.

Okay, let's go.

Okay, so. Okay, now, now, dude, turn it up to 12. Okay, so, dude, TLDR, the board has been been systematically screwing up intel for the last 10 years. This, I think, comes from the history of intel being the greatest semiconductor company alive of the 2010s. Dude, the board was filled with like politicians and like ex senator, what you call it, Secretary of State and like generals and like never anyone in the semiconductor industry. Intel had a like, intel had like a view of arrogance. Like we are the best and you're going to suck it because we're the best. Like you know, effectively. It is an intel first worldview that like has slowly been crumbling over time. So pretty much on the technology side, they've never ever had anyone who actually knew how to run a semiconductor company except for professors or people on Intel. So like you can look back in like the 2003, 2004 era, it's usually two or three people from intel and then the rest of it is just like randos. So they never thought, hey, why do we need a semiconductor person on the board? This slowly became a melting frog issue as they like missed. The 10 nanometer debacle is like the real big, you know, change point in Intel.

And really quickly, can you unpack the 10 nanometer debacle a little bit?

Okay, so 10 nanometer was Intel's process at the time. It was supposed to be the next one after 14. And I think it's plus plus, plus plus or maybe so pretty much it got delayed. So they tried to do a lot of aggressive technologies. They did quad pattern dv, copper, cobalt, interconnects, a lot of things they shoved in here to make a good product, but they just missed over and over and over again. So it was a ginormous fumble. And at the time the CEO was a cfo, so the finance guy over oversaw the technology essentially implode and slowly degrade. So 10 nanometer was just like the true, slow, like you know, multiple years in the making. Train wreck.

Yeah. And what was, what was 10 nanometer like critical for? Is this like mobile transition? Just CPUs.

Just CPUs. Remember this is, this is an intel that missed and totally like was irrelevant for mobile anyways. This is just like data center CPUs, normal PC CPUs. This is when AMD essentially caught up

because yeah, the Ryzen stuff was happening. Thread rippers and whatnot.

Yep. All that stuff is because they're using the TSMC process versus intel using their process. Their process lost the TSMC and then also AMD's better design, whatever, TLDR. Total, total problem. And half of the people who are on the board still are from that era. Like you would argue, like, hey, capitalism works. You should fire people who do shitty things at their job. They should all be fired. And so most of them, most of Them got fired. And then I think this is in the semi analysis piece. But essentially like the people who remain effectively was the guy who just stepped down. I still think he's on the board. And now the chairman is this guy named Frank Erie. Frank erie joined in 2009. So he is just as, as like guilty of anyone. And he's a banker dude. Like that's, that's his background. He's a deals guy. So the first thing you need to know is a deals guy when it becomes chairman of a company is going to do deals. So. So that's what he's been doing. He's sold Altera.

And you ship your org chart.

Yeah, yeah, yeah. Ship it, sell it to anyone who can do it. And he wanted to sell the foundry business.

Okay.

That's like the Wall Street Journal thing. And in my opinion total mistake. We think it's a total mistake at semi analysis. But then on top of that like lip Bhutan. I understand.

So really quickly clarify that because I hear deals guy comes in, investment banker comes in, a lot of people rumbling independently for the last decade split up Intel. It sounds like they were trying to do that. Give me like your evolution of like should intel have split up at some point? Should intel like split up now? What is the current stance?

So I think at this point in time I think they probably shouldn't split split up just because like they don't

have a customer, right?

Yeah, they don't have a customer. It's like the best time was definitely like three years ago at this point in time. It's like you need the money from design now just to keep the lights on. And I mean like not even just to have the customer, just to like fund anything, to have any cash from Ops. An ideal perfect world is somehow Hock Tan, the Sith lord of private equity buys the design business of intel and, and scrapes it and guts it to the floor. That has always been a long time rumor that Hawk wanted to do that. Hawk doesn't want to do it because Intel's so bad. So there is no buyer. Design's like really screwed because it's like stuck on this process. Plus it's also filled with bloat. And at this point in time I think the only strategic part of the business that really matters is Foundry. Like let's just like, you know, because outside of the x86 CPU like they're getting their face kicked in by, by ARM.

Okay.

So it's you know, x86 and then they have to compete against AMD, which is still kicking their faces in with TSMC. And so like what is a secular decliner x86 worth? Not much to anyone. And I think the reality is it's like, you know, with ARM and these custom CPUs effectively they are just one of many like competitive products. And I think their competitive edge like dwindles pretty much. They just have distribution into OEMs. So where is a business that has value that matters At Intel? It's a foundry. There's one foundry in the entire world right now. It's tsmc, it's monopoly. And that's the only thing that's worth anything. And I think there's a chance to have a second foundry and that could be Intel. But here we are.

What about Samsung? I mean Elon did that deal. It seems like Samsung's like at least in the conversation as a potential foundry for specific things. Is, is that like, are they really like several orders of magnitude below tsmc?

Yeah, I think Samsung's just as screwed as intel, but it's like a second quiet more screwed that I think is like, I mean the PPA there, meaning power performance area, is pretty bad. It's probably just as bad as intel. But the one difference is Samsung has had external customers and probably could have the external customers again if, if I had to bet between like intel in theory has a better process for sure. But it's like, it's like a bed of like we're going to say horse and saddle. Okay. Like the intel saddle is like you know, six rags like tied together around the horse that only like one writer in the entire world knows how to use, which is intel design on a presumably decent horse. It's like 18A is probably like equivalent to 3 nanometer, obviously not the best in the world, but like that it deserves to exist versus like Samsung's horse is just like really bad. The PPA is total trash. But hey, it has a saddle, people have ridden it before. So there's in theory a customer. And so Liputan is the perfect guy because he ran Cadence, which is an EDA company. And eda, their whole thing is they design. And so once you have like a design you can like put it in the foundry. And so like that important critical step, the PDK is like the missing piece for intel and Liputin could be the guy. But intel sucks.

So yeah, so I mean my, my read on Trump saying we need Lip Bhutan out is like a complete misunderstanding of the importance of intel. And basically he's operating on like a 20 year old world Model thinking that, oh, I'm familiar with Intel. I know the brand intel inside side. Like that's an important company. We need. It's a critically important American company. We need an American at the helm. When in fact it's like intel doesn't really matter. There's like a wind down in process. Let loop, let Lip Bhutan like run it. Even if he's exfiltrating everything to China, it doesn't matter because Huawei and SMIC and SME are light years ahead of Intel.

Yeah.

So I think that that is sort of true. But I also, I will give the Trump admin props in terms of like, I do think they know and understand the critical importance of Intel.

What is the critical importance? What is the critical importance?

They are the only domestic semiconductor process.

Okay.

So only our last.

Our last hope. It's that or release the future from tsmc. And that's pretty much like probably the most likely option, at least.

Future future.

Least the future.

Least the future. Yeah. Because you could buy a process or force TSMC to make a process in the United States, which is what they're doing.

Yep. With Arizona, right?

Yeah, in Arizona. And then possibly more expansions. I'm sure there's some kind of deal. But after you have that like all the R and D happens in Taiwan. So effectively, like you know, the Taiwan constraint thing, Same problem we had before because like a missile flies, boom. We don't know how to make new chips. We would just have old chips. It's like leasing a car. We can't buy the new model. Yeah. So that's a problem. That's like the same problem but maybe more kick the can down the road if we had all the capacity here, but no R and D. Yeah.

So yeah. What is the like moonshot to make intel dominant again or catch up. People were noodling on like, should Elon come in and do the Elon thing and hire the most cracked engineering team and. And drive everyone 10 times harder. But it feels like people kind of kick the tires on that. There was the rumor of like all the PJs in the same at Mar a Lago, but kind of everyone was at Mar a Lago at the same time. Is there, is there a world where like intel needs to go more founder mode? Like Lipouton feels like the epitome of like manager mode. Like would you be like excited about something like that? Somebody who says like, I'm gonna take it private. Be extremely risky. There's a 10% chance that this thing works, 90% chance we completely destroy everything. But if it works, it'll be amazing.

Dylan Battellan is CEO.

Dylan and dog get both of them.

Dude, if Dylan is the CEO, then my life is going to become miserable. Dog, like, look. So I think I personally believe that Fab and Found Fabless and Foundry need to have separate lives because you're like, there are two really drunk adults at the bar tied together hip to hip, and it's like one of them could be dying. Actually, one could be a corpse and the other has like a shot. Okay. So like, I just don't think, I think, you know, like, I believe in focus. I really do. I believe that small teams with a lot of focus can make extraordinary results. I'm sure you see this in technology over and over and over again. I think Foundry needs to be like given a leash that's long enough to make the shot.

True.

Probably a smaller dream and probably a pot of gold at the end of the rainbow. But then essentially be told, hey, here's your purse budget, here's your potential capacity. If you win XYZ whatever, you'll get a ginormous order from in the fabless semiconductor companies in America. And that's the pot of gold at the end of the tunnel. And then behind you is death. And there's only one way and it's forward. Because how I think we've gotten to this point is there was always this second sloppy option of, well, we'll just put the Intel CPUs in there. That's how Foundry has been treated for so long. And I think Intel CPUs over a long enough period of time is just not going to be enough to fill the foundry. So you're going to have to do. You have to separate them. And I think you have to make like capitalism work. Which means that you need like a pot of gold at the end of the rainbow, a true like shot to do it. And then like probably founder mode, if it makes sense. Like, sure. Maybe Liquidan is the, Like Liquidan has much more of the qualifications. It can't be someone who's from Intel. And I think the company should be private because public markets would kill the shit out of it. Like, it's just going to be miserable. It's going to be like a face beating the entire time. And yeah, dude, if Elon did it, I'll be very stoked. I'm not going to lie to you. Like, however you feel about Elon's companies, like, you know, someone said he makes the late. Oh no, he makes the impossible late Right. He, he will. Maybe he'll be late, but it'll be like he'll get it done. Right. There's, there's no one else who's really done this. Pretty much no one's fallen off the leading edge and come back. And so this is like a moonshot. This is a moonshop's problem. And I don't.

When the Twitter deal was happening, I was saying like, okay, we got, we marshaled $44 billion in private capital. There's also the Chips act going on. You package all that money together and you give Elon intel. And like what is that? What does that counterfactual look like in the course of history?

I think the question is, does the stock need to drop another 50% before it can become a viable like actual take private Target?

It's an 87 million dollar company today.

I think you can do it. I think you can do it today.

Yeah.

I mean if you can take Twitter Private At 44, you can take Intel Private 87.

I think, I think you just have to split the foundry fab.

Yeah, it's that simple.

Because the foundry, the, the sorry, sorry. Not Foundry fab. Sorry. Fabless.

Fabulous.

Yeah. The Fabless business is worth something.

Okay.

Like it is in my opinion. The like, dude, Hawk Chan, the Sith evil like PE overlord would like crush it. Like I really do believe that. My favorite.

So quickly like you split out design. You have this Fabless semiconductor company similar to it would compete with Nvidia in some ways when they go into.

It'd be amd.

Amd. And so they'd be focused on CPU and their customers would be who Exactly.

I think you'd focus.

Electronics.

Electronics companies.

But yeah, like I think you would focus. So the thing they keep saying is AI at the edge.

Yep.

I'm a total hater if I'm be honest with you. Like, you know the Internet works packets are pretty quick.

Yep.

What I think would be the best way is they do have some networking content. They do have the PC business. They do have like high end data centers CPUs. Maybe like the high end data center CPUs. Like, I mean to be clear, they're gonna be like a third place or second place against amd. Like it's gonna be an ugly world. But like what you could do is you can just kill all the SKUs and all the expansion you've done over time. Like intel has all these custom SKUs for whatever you make. Like good, better, best data center and I don't know, edge or mobile Some kind of optimized thing and you fit all your products into those categories, you kill all the unprofitable ones, you fire 50% of the people and then you like, you know, do your best to extract the rent in places to that you cannot be ripped out. Maybe you, I don't know, monetize your CPU software, sell it or some shit to amd. Like do your absolute best. And it's like a really sad ending to intel, the fabless business, but I think it has one that is worth more than $0 is.

Yeah, talk about the GPU CPU split here because Intel's never really been a player in gpu. And that feels like when we talk about the value of TSMC in Taiwan and AI, we're talking about super intelligence and these mega clusters and the ability to train frontier models. And it feels like like even, even a high performing intel like isn't a player in that world or, or should we be thinking about it in that grand of terms?

No, I don't think they're a player in that world. I really like it. I think it sucks. But like you know, Gotti is, Gotti is a chip that sits in warehouses around the United States. You know the battle mage CP or GPU is like not the worst product, sure. But I think if you think about just like semiconductors, man, one of my favorite analogies or ways to look at it is it usually ends up being like let's say a 60, 30, 10 market and the 10% like 60% makes 2x the profit of the 30 and then the 10 is like break even or loses money. I feel like Intel's market positioning kind of puts them at the 10%. And I think like I, like I really think the private equity outcome for the fabless side is the best outcome possible. Because like and, and the reason why I say Octan, like let's actually just like play the Hawk Tan playbook. What Hock Tan does is he takes a business that's extremely mature and sticky that was like a winner of the last cycle and then just like fast forward 10 years of the maturity all the way to today. So like VMware is a perfect example. Virtualization was like, you know the thing in the 2010s and made CPUs better and all this stuff. And you know there were this like almost monopoly. They still have the vast majority share but they were like spending all this stuff growing expenses. And then Hawk was like dude, no, they actually have a business that has some terminal value. They're not living in this paradigm. So what I'm going to do is I'm just going to fire half the people, raise the prices on the, like, raise them massively, lose a lot of our customers along the way, but become a massive cash cow and then like essentially squeeze that to the end. But if you think about what Koch is actually doing is he's accelerating 10 years of industry progress into three. Right. Like that's what would happen. But like it's going to be a lot uglier and slower show in public markets versus just like time collapsing price raise, fire all the people. Boom, you're profitable, you're at the end state, you're. You're essentially underwriting no growth. I think intel needs to underwrite no growth. And that's not in the fabless business.

Yeah, it's just embracing reality. Like embracing, embracing the reality and taking, taking your medicine.

Someone, someone in the chat, Sharon says please ask Doug about the new intel factory in his hometown of New Albany, which, what the status is. Do you have any idea what's going on there?

New Albany, not my hometown.

I haven't. No, no, no, not your hometown.

This listener, this listener's hometown is maybe looking for a job in the fab.

So the New Albany, as far as I understand, the ground has broken, the shell is empty and I don't think they'll fill it.

Okay, this won't happen.

That is. But the thing is I think that that's a really valuable asset that needs to be remarketed. I think, I think it's probably likely someone will buy it. And if I had to guess, it's like tsmc. So. Yeah, I don't know what that looks like, but I think, I think a really shrunken down intel looks like just Oregon and just New Mexico and Arizona. I think so, yeah. Arizona and that's it. The capacity that was expanded for Ohio was like a YOLO bet the farm somehow where like be meeting AMD again. Like I think it was underwritten to have this like. And that's the reason why Lip Bhutan was like, dude, touch grass. This isn't happening. Pack yellsinger. This is just like, this is not happening. Like you're like completely not realistic. I think that, that it just doesn't like the capacity they had there for the. The wafer starts is just like so large. It doesn't make any sense. Yeah, and I mean like dude, the Lip Bhutan thing and like, okay, how do you feel about like the, you know, Chinese thing? I mean there's just not many people who could take, take the job and I mean it really sucks and it's very hard. But like, dude, he actually embraced reality. Like the most realistic thing I saw, he's like, dude, we're not even a top 10 player in AI. And I was like, yeah, that sounds like reality. Like this is the first realistic thing I've heard for a long time. And that's like refreshing to me. The guide was totally like the, the print was totally messed up, like, you know, effectively. It was really ugly. Essentially they're like, yeah, we did pull forward tariffs. Yeah, the guide is a little messed up stuff and yeah, it's going to be uglier from here. But I kind of vibed with it because it's like we're just going to tell you like it is like, like how it is. You can like trade the expectations of this but like the, the stock and company and the value of it is in my opinion a binary option is what is the foundry worth to America and how do we get, how do we like fast forward this like painful capital intense world where we like don't have a leading edge semiconductor producer that can have an external customer to one that can. That's, that's the entire value of the company to me. And then Fabless is like a call option cash gusher that can be hopefully harvested.

Strategy is we need people to rip the band aid off.

Yeah, yeah, yeah. It's going to cost a lot.

Give me the update on GPT5.

Yeah, specifically it. This was the week that the timeline woke up to language models plateauing. A lot of people are bummed about it. Bearish. I think in it was maybe six

months ago that Dylan Patel went on Dwarf Cash and said like, if GPT5 is good, we're totally good on all the semiconductor spend. If it's not, maybe we're in trouble. And so I'm wondering like, if there's updated thinking around there based on, you know, do we, do we need any more inference clusters or training clusters or is it just a game of where is the profitable inference happening? Let's make sure we have enough to meet demand but not go further, not create some overhang. How much of a dance do we have to do to make sure that we don't get over capitalized, over built?

Okay, so this is a pretty interesting question. I'm gonna parrot Dylan on Twitter because, you know, like my lord and savior Dylan knows, knows what he's talking about. GPT5 is a little disappointing, I think for the power users of Twitter. There's no other way. There's like, no other way to put it is that, like, for the power users of Twitter who've been like, chugging O3 deep research things for a long time, and it's like messing around with rock and like, messing around with whatever. I personally do not feel like a massive difference. It's less verbose. Whatever. Maybe it's slightly better. I don't know if you're seeing this thing about, like, the, the SWE benchmark, like, app. Like the. So, like, the, like is. Yeah, the chart. No, no, not the chart. The SWE benchmark. If you look at it, not all 500 tasks are measured. They pulled some 477. It's like 477. Yes. Yeah, yeah. So like, that's a, like, that's an apples to oranges comparison. So that's like a big deal. But I do think you have to think about, like, okay, so on one hand the cultists and like, me and like the power users are probably disappointed. Disappointed. On the other hand, you know, there's 600 million free users who probably just woke up to like a big upgrade. And I think that that's how I feel. Like, this was actually, like, that's the actual strategy. They pushed them to five. Five's a lot better versus four for, you know, 4O. And it just. Sorry, dude. The fucking naming.

Well, I think. I think another way to put it is if you're a power user around, like using it for research and learning, you're. You're maybe disappointed if you were a companionship power user. I mean, we were reviewing. I mean, r.chatgpt on Reddit is absolutely in shambles today, dude.

I know.

Because the Girlfriends RP rip. Well, there's a new entrant there, I think. I think they might have to peel off that market and give it to Elon.

Yeah, that's true. You can do. You can do the Annie Network. You know, dude, I don't. I, I. Yeah, I think that that's the correct way to think about it. I think it feels like it's. Try to be as simple as a. Like a simplification of the. Of the SKUs, because as you guys know, it's confusing as hell. There's many, there's like many thinking and, and flagship for the most part. I had to, if I had to guess, flagship probably rent, like, probably prints cash for them and is like, much cheaper and like a smaller model. Maybe not a smaller model, but like, compared to like a very big model smell. Right. It's not big model smell. So it's Very. And then also if you look on the pricing, it's pretty cheap and if it does what it does well then just like we talked about, Swede benchmark might not be, might not be correct but like you nuke the price on your competitors in terms of like Claude versus chat GPT5 tokens. And so I think that that's like the, this is the better, bigger, faster or maybe it's the faster, cheaper but not better update, if that makes sense.

Yeah, well one thing that stood out yesterday, so they have 700 million weekly actives globally, 85% are outside of the United States. And you can just imagine in that, in those hundreds of millions that are outside of the U.S. many of you know like the question just becomes like how they can continue to serve the like will they be able to serve

what will ARPU be and what will the margins be on the, on the incremental International Developing Markets ChatGPT user because the marginal cost to serve WhatsApp or Facebook even, even yes you're storing images for, for Instagram but or, or WhatsApp you're running a database but the marginal cost is so, so low. And it feels like at least right now, maybe it changes in a couple years but right now the marginal cost of serving a ChatGPT user, even a casual weekly active user is, is in the, it's in the dollars per year. And when you look at the monetization rates of Facebook and WhatsApp in developing markets, the ARPU is like dollars per year. And so matching that up feels like a big challenge. And I'm wondering about yeah like ASICS AI on the edge dedicated, you know, servers or, or even more pared down models. What's the solution so that you're not just burning a ton of cash supporting unprofitable users.

I wonder if we'll have the bandwidth to get this done if it makes it to a big post. But I think we're thinking a lot about this deeply at semi analysis I think the cost of a free user for a small model is small, surprisingly low. I want to be a little bit more sure about that. But I think if you have high batch size with you know, a B200 or a GB200, I don't know if it's dollars per year. I really, I really think the actual net tokens that you're, you're, you're able to do is a lot higher than people think. And I think that there could be this as. But the problem is, is like the same, there's the same issue, Right. Think of it as like a pareto curve. Right. 20% of the people are paying at the wazoo and more than happy to pay for like a crap ton. Right. The pro and plus of OpenAI. The only problem is like the, the, the top 5% of that 20% are paying or like, you know, they're massively losing money on. And then you have this like long tail that is like kind of interested, likes tokens becoming a little bit more addicted, easy, cheapish to whatever. And then as you can shift that up, maybe you can make the revenue that you're not, you're losing for the hyper, you know, for the pros. Right. And so that's kind of the, I think that that's kind of like the more interesting like thought process. I do think the unit economics of these things are a little bit like, I mean it really depends on the size of the model and then like what inference they're. They're doing. But like my impression is you can be doing, you know, bill, you know, millions and millions and millions of tokens for relatively cheap a year. And if you monetize, if you know we're talking about like you can monetize like 15 to 20% or even 20 to 30% of your, your tokens are actually being monetized at like, you know, the market rate, you can pay off the 80%. And that break even I think is a really important like question. I don't know the answer to that. I know we're working on it, but I don't have anything.

Yeah, I think the, I mean the, the going back to your like Instagram or WhatsApp comparison. Instagram and WhatsApp are pretty much the same wherever you are in the world, regardless of how much revenue you're generating for these platforms. But it's very possible in the future that companies cannot serve the quality of model in some of these areas that they can.

Well, they will. It'll just be $200. And $200 in an emerging market is like a ton of money. And so the adoption rate and the upsell rate will just be way, way lower. There'll probably be like an ad supported

tier, but again ads themselves don't generate.

The ARPU will be much lower in emerging market markets.

Yeah, and I think that's happening. I can't name the hire, but I know the hire that essentially started the ads.

Yeah. Fiji platform.

Yeah. Is there. So like, you know, that's the freaking surprise. I wouldn't be surprised. The answer if there's Like a Twitter user. The answer is always ads. Okay, there's gonna be ads eventually.

Yep. But they probably monetize, like, within an order of magnitude of how well Facebook ads monetize and Google Ads monetize. So that's fine as long as your inference cost is roughly the same as within an order of magnitude of Google's service cost and Facebook's service cost to serve a user, which are pennies, I imagine. But it does feel like we can get there.

I think the timeline broadly woke up this week that we're building developer tools and consumer Internet companies, and that's the. The game here. And machine God is, like, real in some ways, delayed. I'm curious, how do you think. How much do you think the labs care about the student market? There had been some reporting around the token generation. As school ended last year, token generation dropped dramatically. And then today, Sundar came out and shared that Gemini, they're making Gemini like, you know, like unlimited, sort of like free for students this coming year. And in some ways seems like he's

staying in the game. He's like, I'm not letting chatgpt.

Well, no. And it feels like that's potentially a way for them to like, to kind of corner users that will become very valuable, maybe aren't so valuable in the short term, but will become very valuable over time if you can get them to stick.

Yeah, yeah.

I think let's kind of talk about this because I think that's like, just a really interesting. And I'll kind of go into Google. I think, you know how Cuda, you like just analogy, right. How Cuda became a thing is because Jensen gave it away to PhD people for like a decade. And then all of a sudden, everyone who ever did anything realistically in machine learning was like, well, I'm using Cuda because that's what I've been using. I think it's important because if you're thinking about, like, you know, where are the users of tomorrow? The S curves of penetration are like, it's not going to be the old people who are using the hell out of it first. Who knows, maybe they are. But it's like the young people are probably going to over index, you know, people my age, which makes me feel old, is like going to, you know, probably index about. Right. And then people older than me are going to index under. And so where do you actually win market share today? It's probably not in the people my age and older or the power users or whatever. Whatever. You probably need to, like, get a. You Know you need to get a 12 year old addicted to Gemini which sounds like very terrible now that I say out loud but you wouldn't say

that about Google search, you know or Google Drive or Gmail. But yes it is, it's low churn. You have to get them over the adoption curve so that they are low

churn and that's where they're like, like I think it's like a cognitive referent. Right. Like right now even though I am very model curious, meaning that like I have a, I have you know a paid sub to like Claude to grok to to ChatGPT recently especially like you know in the, you know the good old golden era days of O3, you know, 03 Pro deep research. I'm just pounding deep research queries and like agent queries constantly of course but like and I'm going to be, to be clear, I'm going to start looking around a little bit. But like I think, I think that that's what Gemini's like real issue is is they need to become the cognitive referent and I think when people open up a you know, a GPT like trans, you know, transformer model effectively to ask them in LLMs they need to be thinking of Gemini in the conversation and they just don't. Dude, sometimes I like talk about LLMs with like people and it'd be like you know, OpenAI and anthropic and what's the third? Oh yeah, Gemini, you know the third. Yeah. I like weirdly forget about them and I think that that's the, they don't want to be forgotten and I also think if we do a segue too, I think that's why they're also strong starting to sell TPUs externally. This means like as a service, right in their cloud, right. They're selling like they talked about on the last call where like you know, GCP is starting to accelerate and I think where we're starting to see that is I think TPU is starting to pick up a little bit.

Do you think that'll be like distilled models that are fine tuned or developed for TPUs specifically or something else?

I think it's, I think it's more that like there's a few model companies who are very interested in actually capacity. You know, for example Anthropic is definitely, you know, definitely uses some TPAs. I know that for sure. And so you know in the framework

they use Trainium a little bit too.

They also do use Trainium. I think it's very funny because Anthropic uses everything.

Interesting.

Yeah.

Is that like a huge cost center to have kind of like a replatforming team to rewrite all your CUDA code into Trainium and TPU compatible code?

I don't know, probably. But Anthropic is cracked, so that's what they're.

They have Claude code, they have full access to CLAUDE code.

Re platforming. I mean that is the best thing.

They don't have to worry about being dropped.

Yeah, I mean I was wondering like when will a cross compiler exist and you can, I mean, didn't Facebook do that for a while where like they would write JavaScript and it would compile like C or something like that because they'd written like so much JavaScript that they were just like, okay, we don't have time to rewrite all this, let's just write our own compiler. And. And you can imagine Anthropic doing something similar to run on TPU or Trainium.

Yeah, I think that that's happening to a certain extent. I think that for Anthropic they think of it as optionality and I think Trainium specifically is like, you know, the, the other drunk suitor at the bar who like really needs a partner like Amazon who's now last place, right?

Yep.

Last era was first place, now they're last place. And then Anthropic who is like, you know, the scrappy number three who needs capital compute is like, come on dude, marriage made in heaven. Let's tie it up together. And then that's why training is just like ramping.

We gotta start using these drunk bar analogies.

I love drunk bar analogies.

We usually stick to the sports analogies, but I'm probably more familiar with the drunk bar analogies. Those are good.

Ross McCannell in the chat says, please ask Doug about the state of memory makers. While sk, Hynix and or Micron be able to maintain margins and keep. Keep HBM as a somewhat differentiated product or is it doomed to be a commodity business?

I think this is a good question trying to think about how to phrase this. So I think one of the ways to think about this first and foremost is HVM is like the new dram. I mean like there's like a weird, weird aspect where it's like, okay, DRAM was like the best and you can look at the cycle's margins and like they're not actually higher than last cycle's margins, which is like probably a little bit of a. I wrote about this like being the super new dram. I think HBM is really, really really important. Super, super duper in demand and is like you know the new memory but in the other ways NAND specifically is kind of like the old memory and it's getting much worse with the commoditization and the like especially CXMT on the lowest end or no YMTC on the lowest end and then CXMT on dram. So you're starting to see some, some of the aspects there get a little worse. But I do think NAND is trying really hard to kind of hold on to the you know the oligopoly where everyone doesn't raise bits together and maybe who knows it becomes so old and whatever you end up like HDDs which are actually having like, like you should go look at STX and what's the other and WDC because they're like their stocks are ripping because yeah Western Digital, yeah Western Digital because they're not investing. They you know it's a two player market. They even gave hammer to the other competitors so that they can like effectively like you know have a two player oligopoly where they control supply and this

is just, this is just the Nash equilibrium. This is just like game theoretic. They know that they, that this is better for both of them. So they're not getting a price war right?

Yeah, yeah, yeah, 100% but that's only for hard drives. Like NAND isn't there yet man you have like a slippy Chinese slippery Chinese like swing producer that is willing to blow it up.

Got it.

And so I, I think on HBM I think it's kind of complicated. I don't want to like say too much because I don't even know, I don't even know what semi analysis is. House like house house view is right now let's just like I'll talk about some facts I think might be very interesting and just like thoughts is that like look last time this year Micron and SK Hynix were effectively completely booked out and obviously Samsung wasn't even in the game this year they're not booked out. What is the difference is I think it's the threat of the Samsung quality qualification. I don't, I don't know about the Samsung qualification. Okay. I don't think anyone knows the I hear out of Korea like I hear that they're qualified, they're not qualified. They're qualified, they're not qualified, they're qualified but you know they have to hit some kind of yield thing. I hear kinds of all the time. I have no idea but I think the fact that HBM4 pricing seems to be a little bit like dampened just on the threat of qualification really kind of tells me that the HBM cycle is really long in the tooth and is probably where we're nearing the beginning of the end just mathematically as the second derivative continues to go down for most of the memory stocks you can look at like SK and Micron. I still like sk. I still think what they do is really, really differentiated and valuable on a relative basis. I mean they got to HBM first, they still are really awesome, they still have the best, the best product and process but I guess I'm just kind of like you know, I'm a cycles brain guy and it's been about three years of a tight cycle in memory and you're like you're asking me to bet on the 4th. I know what the base rate should be you shouldn't bet on but I don't think we actually have a house view other than that like because at the same time you know, Nvidia goes burr and HBM bits goes up and even though there's a little bit of an oversupply in HBM3 I still think HBM4 is going to be pretty good but you know, like rate of change brain tells me it's a little spicy.

Where else in the supply chain should we update? If we're totally plateau pelled and we say we're not going to necessarily even want to do the next order of magnitude pre training run the mega cluster, we're maybe going to be more distributed, more inference heavy. Are we thinking ASICS? Are we thinking more focus on depreciated GPUs, distilling models to older hardware or just you know, really getting all of the juice out of the current H1 hundreds and, and GB2 hundreds and like actually just like you know like not, not re platforming to the latest and greatest constantly like not, not worrying about that or like where else, where else should we be kind of updating on various parts of the supply chain?

Okay, so I, I think a lot about this or in one I feel like I'm like the wrong person or I'm probably a decent person to answer this but like it's going to be really speculative so I'm sure like some massive disclaimers who the F knows, right? Like maybe you know, maybe there's one more thing or Claude comes out and they cook on five or something, I don't know or Gemini next week makes us all pilled. I don't know.

Yeah.

5 I think we're. If we're okay. So let's just like walk through some of them. The massive pre training cluster doesn't happen. I think the fiber from data center to data center is like the most screwed. That is the most whipped at the tail. The multi data center training thing.

Sure.

That is like ouch, goodbye. You don't need anymore. Right. You can do 100k clusters of RL and you get, you get. And so that's probably where I think if you had to like you know, slice the puzzle where it's most impacted. It's probably there for training specifically. I still think we would end up doing a lot of inference. I do believe, I really do believe that if all progress stops today we probably would still have a decade of productivity as this technology is ingested. It's not AGI God but it does happen to be an amazing densification simplification densification of all the information into an answer machine.

Fucking awesome.

Useful.

I mean it's like, it's like the rollout of CPUs in the cloud. Like we're going to use CPUs in the cloud for a long time. Took 20 years to actually do all the things that you can do with CPUs and databases and hard drives in the cloud. And now we have GPU and lm. So we're going to, we're going to stuff that in every single cranny of the economy.

What did you think there was a headline recently that Microsoft, it was Azure's non AI cloud business was growing at home Almost.

Yeah.

I think, I think the final read on that was that that was just driven by OpenAI as an Azure client. But basically the read off of earnings was that their core infrastructure group grew faster than their AI services group. So more people buying compute as opposed to buying tokens. But that's not necessarily a read that like people are spinning up more, more cpu. But walk us through it actually that's a great question.

I'm very familiar with what you're talking about. The Azure beat last quarter specifically was driven by the infrastructure side more so than the token side. And then ironically one of my colleagues at Semi Analysis pointed this out. There wasn't the disclosure they had last quarter. They disclosed what the percentage AI was to the growth and they did not talk about it. And if they stopped disclosing, I can tell you the answer is it was lower. So I think that's probably a pretty interesting thing. It tells you that the consumption is definitely worse I don't know how they're classifying.

Well, it doesn't really. That tracks with Satya canceling data center developments and just saying, I'm happy to lease. Also the business on a tear, doing multiple rounds. They've done more layoffs this year than I think the last, like, three years combined. Which tells me that everything that, like, the timeline is waking up to this week, like Satya, has now been pulled on. Even if you look back at some of his interviews, he's talking about, like, yeah, OpenAI is like a great consumer technology company.

We're happy to partner with them. But wait, wait, sorry, can you. Can you restate and unpack that a little bit more? So there were. So Azure grew a ton. AI tokens grew a ton, but AI core infrastructure grew even more. But they used to break out within core infrastructure. How much of that infrastructure is being used for AI versus not. And they stopped disclosing that. Is that correct?

Yes. Okay, that's correct.

Yeah.

And so by not disclosing it, you know, the answer is it's slower.

Yeah, but it was still. It was still like, pretty big growth overall. So it's just that the. The second derivative. Slow down. Yeah, still accelerating, but it's at a lower rate.

Yeah.

Okay, interesting.

Well, actually, I feel like now I'm talking out of my ass. I don't see the number in front of me, but, like, look, I think the thing, the Microsoft print is like, they pretty much are like almost mid cycle, pressing the break, and then they, like, you know, they juice. But, like, you know, as it decelerates the investments, you still get to reap all the investments you made.

And so they have a massive backlog. Right. Don't they have, like, $100 billion of cloud backlog?

Yeah, and so all this backlog has started to become revenue, and then Microsoft still is such a winner. Fast or rewind to three years ago, when Microsoft was clear number two, and now they're outgrowing on a much higher base than AWS is. And AWS is not in this. Right. So I think that if you think of it from the perspective of, hey, Microsoft Azure versus versus aws, did you know? Did Satya win? The answer is yes. But I think the thing that is interesting definitely going forward is, like, how this all works and how OpenAI continues to finance and spend and pay for more compute infrastructure. Because at the same time, one of the other ways you could look at it is the entire math of where this is going is just like, all at Oracle's Pocket. Oracle Effectively. So you talk about the Microsoft slowdown. Oracle literally just completely ramped up when Microsoft slowed down. And then, and then Oracle went to OpenAI. And OpenAI is like, great, dude, we got a new customer. They're willing to do Stargate. There was even a spot that was supposed to be Stargate for Microsoft that kind of like, you know, and then

all of a sudden he said, I'm good for my, I'm good for my. However many billion.

80 billion. Yeah, but then, but then, but then that, that deal went to Oracle, right?

Yes, that deal went to Oracle. Okay, so I think that that's. Yeah, I think Microsoft is like slowing down. They definitely had this massive lead, like truly a massively. Remember like last year, the perception of who is winning the race for the hyperscalers. It was Microsoft number one, because the OpenAI partnership plus the infrastructure, you're like, dang, they're so far ahead now, being conservative, kind of pulling off, you know, decelerating. But I also want to put some like, let's also talk about this a little bit because I think the underappreciated part about the Microsoft thing is, yeah, just like how shitty they've done with access to the model this entire time. They've had access to the model weights at OpenAI the entire time. And you can also argue that no company has more competitive threat for, for their core business than Microsoft does. I don't know about you, but I do a lot more editing and drafting and like information search and ideation in ChatGPT than I used to do it in work. Right. So you're like typing up something and

you're like, yeah, so. So. So Elon posted yesterday that Microsoft is get. Is going to get absolutely cooked by OpenAI. And in many ways, obviously, like, Microsoft has a ton of users for copilot. They're taking AI seriously. But what you're getting at is, is one, you know, and it's hard to read too much into what Elon is saying because, like, he's got a lot of, you know, he's playing games behind the scenes that we don't necessarily have a view into. But what you're getting at is interesting, which is you might. Instead of there's a world in the future where instead of opening a Doc or an Excel sheet, you just start talking with the model and saying, I want to make a model for this product line, you know, going out over the next like three years.

It basically has Microsoft 365 as a tool that it can call and then eventually instantiate A Word doc for you if it needs to.

It creates the output in, in, in,

in there in the actual application that they have because they already have Cloud hosted PowerPoint and they just are like, oh, it sounds like you're going down a tree. You need a PowerPoint. Here you go. I generated it and you can edit it if you want. But also here's the export. We got a PDF for you here because we used our tool.

Yeah, 100%. Yeah, 100%.

And like, but yeah, huge, huge UX problem. Like, like, I mean like Google has bought, bolted Gemini onto every product and the adoption's been very rocky and it's been tough. It's not as easy as just slap a text box on it and you got a winning consumer product. It does require innovation, I think, in

terms of product development. So going back to the sort of dynamic between Microsoft and Oracle, I think you have to ask the question. If Microsoft is pumping the brakes a little bit and Oracle slamming their foot on the gas, it's going to take a little bit to see who made the right call.

Yeah.

Do we plateau or not? Right. That's the answer.

Well, and I think the answer this week is yes.

Yeah, but I mean, at the same time, like inference for ChatGPT, not necessarily usage wise, could just continue to continue to increase. And if Oracle can position themselves as like the key cloud provider for that, that could be. It's mostly just like, like the Oracle investments just can't be overly aggressive at any point in time because if there's a pullback then they're like, oh, we're unprofitable for a little bit on this.

Yeah, they could take the biggest bath potentially. Yeah, dude, I think. Okay, so I'm trying not to call it because as you know, AI changes. I feel like week to week sometimes it feels just like, I don't know, hyperbolic time chamber. Dude, we could be so over and then it could be so back.

Right.

If you guys remember, pre training is over. It felt very over. And then now we are so back because of essentially the increase in reasoning test time. Interesting. Yeah, sorry, reasoning. And then now I'm thinking about this is like, okay, so the unspoken part about this that makes the ChatGPT5 a lot better than other things is like long task and agentic and tool use. Right. So that is the, in my opinion, and I hear this kind of shit all the time. I thought it's really BSE until I've been messing around with cloud code. It's like, oh, agents, agents, agents, agents, agents, are going to be this big thing and you're like, what the fuck does that mean? Because you're like, I don't know what this agent is doing, but like quadco is pretty cracked, man. You just like ask it to do things and it just like does them and it does it in like kind of a scarily good way. A good example is like we've been hiring people. I think we have some hires in the line. We make them do a case study. That's like not a case study is a great way. Okay? And I was so pissed and annoyed by some of the case study qualities. I made like each of the models do the case study. Dude, you know what did the best case study of them all? Claude code did okay, Claude Code over OpenAI agent. And I think that that's kind of what we're like. Those extremely long context autonomous ability to do stuff on, on its own is going to be the like, you know, the nirvana that like changes everything or makes the consumption a lot bigger. And I think you can kind of see the glimpse of what an agentic feature looks like via Claude code. And then you just assume that instead of them rl ing the shit out of like software and making, you know, the best, but you know, the best commit or something like that, they're going to rl the hell out of like advertising or creating the best, you know, advertising media or they're going to rl the crap out of like finance or making the best financial.

Well, yeah, I mean the thing that stands out to me is like this compounding advantage of like you're a lab, you're competing with other labs and you have access to the, the best coding product and you can use it as much, as much as you want forever. And it gets better and then you guys get better and higher output. And I don't think we've seen this. You know, you didn't see the same dynamic with like Microsoft having a better version of Excel and like. Or the example of like it's not like Mark Zuckerberg with Facebook was like, I get to use Facebook more than my competitors, so I'm going to be better. It's like they actually tried that they

were using Facebook for internal comms. They still did well.

Yeah. But whether or not that gave them

a compounding advantage over teams or.

Yeah, like I think there's this like project or it's like AI 2027 or something. There's like this. I think that that actually feels like a little bit more grounded and then like some of the like, the really like, you know, AI demerism. But I think the like the recursive ability or like whatever, like P. Doom, whatever you want to talk about, but like the recursive ability to make your products better by continuing to invest is like, you know, it is a flywheel. And that's definitely, I think, the anthropic bet that they're going really hard at. And so, you know, the thought process is. It's like, well, okay, if, if we can get the AI agent to do, you know, AI AI experiments, not just coding experiments, then like, boom, we are off to the races. And that really will be, I think the, you know, the, the like where the curve bends back in on itself. Tbd. Right. Yet to be seen. I don't like, I don't exactly like from the outside looking in, there isn't exactly any, any kind of special ability for me to say that it will or will not be that way. But I think so far, from what I understand for the researchers, I don't think everyone's like doomed or beared up on. On some of the stuff. I definitely think ChatGPT5 is a little disappointing. But maybe, maybe OpenAI just isn't cooking like it used to. Right. Like, I think. And I don't know if, I don't know if it's a. Like I don't know what actually happened at 5, but like I want to like, we don't actually know what's the like the ratio. Well, you have to reasoning to like

just totally just remember, you know, you can debate on the people that have left OpenAI in the last couple months, like how good they were. Were they the best people? Were they mercenaries? But any company that has been gearing up for a massive product launch, like, or have gone through a massive product launch, imagine going through that again. But you lost like 40% of, of. Of some of your most elite team members. Like that could be. That's like, should obviously have been a factor here. If it wasn't, then Zuck is cooked because he just hired a bunch of people that just weren't that great.

Someone's correct.

So I think that, I think that, yeah, that's worth asking the question. What would GPT5 have looked like if Ilya was still at the company? What would it have looked like if Mira was still at the company? What would it have looked like if the long tail of researchers that left were still there? What would it look like if they just didn't have the distraction of the talent war? Right.

Yeah, I think That's a valid question because my understanding of why Gemini 2 randomly was so good is Noam Shazir was back. Yeah, that's it. Like, that's literally it. As far as I understand. Like, all of a sudden, Gemini starts cooking again. It's like, yeah, because like the guy who invented half of everything is back. I'm not, I'm not surprised that that's like a real dynamic.

Yeah.

But, yeah, we're gonna have to see. We're gonna have to see. I. I do think you're right vibes are. Are interesting. I do think probably tracking that cohort, like, if you think about it, just like an incremental slosh, like, you know, that's either the biggest, best investment of all time or going to be the worst investment of all time. And like, tracking that cohort and how that works out is going to be like a really interesting case study.

Can you give us an overview of what's happening in private credit headline this week? Obviously that Metta had tapped Blue Al and Pimco for like a $30 billion, you know. You know, I don't know that this sort of pace at which they'll get access to that capital, but just give

us credit 101 and go as deep as you want.

Yeah, 101. But also like, you know, the real kind of risks surrounding it over the next year or so.

Okay. So I do think smeanalysis, I'm the finance guy, which is funny because I definitely don't know if I, like in the big scheme of finance, how much finance I have.

But.

Okay, private credit is like public credit, but there's no marks. Okay. One of the things I got really interested.

Okay.

So one of the things I got really interested about that is like private equity rules because you could suck at your job, but if you have no marks, like, you're, you're, you know, you don't have a bad like performance and your volatility, like, you know, kind of chills the F out. So on an alligator base basis, they're stoked. Like private.

I mean, this is the beauty of being in venture broadly is the market goes up and down and 90% of my net worth is absolutely stable. It's not stable in reality, but yeah.

Yeah. And I think that that's like almost like a feature.

Totally.

It was a feature that was like a bug at the beginning, if it makes sense. Because it's like, okay, there is no mark. And because of that you can effectively, like, you know, the volatility in these assets are really low and you hold it to maturity. There's a lot of studies and papers that effectively like, you know, public equity levered up over five years when you don't sell is like very similar. Like the returns actually start to approach each other. And some of the cohorts of private equity specifically have really started to implode a little bit and have a low money out of the investment. So anyways, where that's just private, that's private equity, which I'm like super familiar with very like very funded private credit is like a lot of the same, same energy in terms of the mark. And effectively you own this piece of debt to maturity. So there's no, you know, you don't really take a mark on it. And, and so you could get like at one point specifically everyone raised these capital vehicles that were like, we're going to get high single digit returns with very little volatility forever. And so everyone was like, dude, sign me the hell up.

And so yeah, the number of billionaires that private credit has created in the last like two decades.

50 in the Forbes article. Tons.

Wait, really?

Oh yeah, yeah, I'll send it to you. Yeah, private credit, I mean just in Aries there's like four billionaires.

Because it's just if you're, if the best way to become a billionaire is Aum Maxing and this and private credit allows you to like Aum Max better than massively anywhere else, almost any of these other sectors then it's just really

Quickly, Bloomberg highlighted 18 folks who are now billionaires from private credit. Starting with Ares at 13 billion net worth going down to blue Al Craig Packer at 1 billion. And so 18 new billionaires have been minted from the private credit boom.

And yeah, I think from my perspective on the private credit boom is like, okay, it's definitely been a thing this entire time but like in the last few years like there is a hockey stick moment where I want to say it's like late 23 or something like that where the pitch was like almost, you know, unbelievable. Effectively.

This is simply too good.

Yeah, yeah, it's like, it's like you can get equity market returns in the long run with, with no mark to market risk and you know, volatility and in theory less risk. So you're like, dude, equity risk or no, no lower than equity return with like relatively no risk on like, you know, it's like obviously spread over Treasury. You're like, bro, sign me the hell up. Like I'm going to slam that button until the button stops.

Working.

And so everyone raised these giant funds, like ginormous funds. And so now these private credit guys are like sitting on a bajillion dollars of aum and they're like, dude, how do I get this, get this to work. And so now private credit is finding its way into, it's kind of like the private equity they're going to have to deploy. And so for people who are looking to deploy very large amounts of assets, I think data centers are going to be really interesting because in theory, and this is just like depends on the whole stack, but data centers are more like real estate investing. So you have a very different return profile that often are baked into the, the deals. And I think it's, and it's very capital intense on a relative basis. So you can deploy a lot of capital, which is awesome. And so you do that, you get like these 5 year, 10 year investments with like pretty solid chances right now at least in like the two, three years. You're super money. You're super money. Good. Like some of the early data center investments are like awesome. And so you're just gonna, you're gonna slam that bit, dude, you're gonna freaking, you're gonna, you're going to deploy.

And so the, the plateau theory feels good here because the risk was that I build a, I build a frontier model capable data center that runs GPT4 and then GPT5 comes out and I can't run GPT5 and I'm completely. All the workloads move to GPT5 and, and I have no business whatsoever. Instead it feels like the workloads are

like sticking around for a while and it's important. You can have like a plateau in intelligence that is different than a plateau in usage.

Exactly.

If usage and demand plateaus, that becomes a real problem. If you just, if you just spent $10 billion on you know, super levered data center development.

Yeah, I think, I think you're probably right in terms of the fact that if the pace of, if the pace of everything slows down in terms of progress, you can underwrite the return, you can underwrite everything a lot more chill. You know, you're not going to be like dude, my, you know, the data set, you know, the chips might get used longer so you can be like, ah, you know, they're four year lives and they're five year lives with high, high certainty. You know, the algorithms aren't just going to like massively consume all this stuff and make like the crap you bought effectively like super cheapened very quickly and Then your investment is more money. Good.

Yeah, like having a bitcoin FPGA farm or something. You got destroyed when everyone went asic. Right. And that's the risk. But that's not the nature of the current frontier path. Like GPT4 workloads or GPT4 class workloads are sort of sticking around probably for a really long time. And in the future, even if we do develop the super intelligent model, it'll probably be calling less intelligent models. There'll be distilled models for specific tasks, specific models for tool usage and all these different agentic workflows, et cetera. So yeah, it feels like all this stuff is going to be sticking around and, and good news for the depreciation cycles.

I, I don't know if that's like. So I'm going to like, please. I'm not going to endorse that view just because we're like speculatively one shotting this in a day. So like, you know, we'll see. I got to think about this. Yeah, but I do think, I mean I do think this is probably better for the longer tail if frontier model stuff slows down. I think if we're just talking about like hey, truly everyone else, right? Like in the world where progress is exponential and there's only three companies doing it effectively, everyone else is a giant fucking loser. Right? Everyone is just a giant fucking loser. There's no way they're going to have any kind of products. Everyone's just like totally cooked. Who cares? Why would you ever invest in Neo Cloud X our front or you know, behind the curve, lab Y or you know, whatever accelerator company Z. Right. I think it probably is better for just the entire ecosystem if we're talking about just like the longer tail of capital specifically like I think like the Neo clouds, the GPUs, the, you know, everyone that isn't named OpenAI, Google or Anthropic. So I do feel very strongly about that.

Well, this is fantastic. The chat is going wild. One hour guest. Semianalysis in my veins.

This is our first ever one hour guest.

No, Dylan Patel was also a one hour guest. Every time I get someone from semi analysis, stay on forever, do the whole show with us. I really enjoy these chats. I really thank you for taking the time.

Just so you know, like we have a lot of, like, I have like a lot of ability to make this happen or not. Dude, we have a lot of cracked people at semi analysis.

Please send them all over. I love these chats.

You talk to Jeremy, right?

You talk to Jeremy, right? Yeah, yeah, he was fantastic.

I fucking love Jeremy. Okay, so like, dude, we have. Yeah. So like that's like the benefit of semi analysis. There's a lot of cracked people. We'll send them all over.

Yeah, send them all over. And I have a plan. So our plan is, we've been, we've been teasing that, you know, basically if you're, if you're a real VC, you're on this like the super secret $1 million a month semianalysis plan. And if you're not on that plan, you're kind of just a tourist. And so we're going to, we're going to get every single VC in Silicon Valley on the $1 million a month semi analysis plan. That's our, that's our pitch. We're going to move it into reality.

I'll take it. I'll take it.

For every billion dollars of a, you should be spending at least 15 million

I think so on. I think so. Otherwise you're just a tourist. You don't really understand this stuff. But thank you so much for stopping by.

We.

Yeah. If you're not, if you're listening to this and you're not subscribed to semi analysis, what are you doing? What are you doing?

Thanks for joining Doug. This was fun.

This is really fun. Let's do this more. We'll talk to you soon.

Yeah, we could talk talk earnings. Honestly, that's, I'm. That's like probably where I'm best, so.

Amazing.

So take care.

All right, amazing.

Bye.

We'll talk to you later. Bye. Up next we have Mitchell Green from Lead Edge Capital coming in the studio in the restream. And while we're bringing him in, let me tell you about ad quick out of home advertising made easy and measurable. Say goodbye to the headaches of out of home advertising. Only Adquick combines technology out of home expertise and data to enable efficient seamless ad buying across the globe. Mitchell, are you there? How are you doing? Good to see you.

How are you?

I'm great.

Great to see you.

Sorry to keep you. Happy Friday. Sorry to keep you waiting. Sorry I missed you when you were on the West Coast. We'll have to hang out in person soon, but how was your trip? How are you doing?

Good, no complaints. How about yourself? How's life?

We're doing great. Big week last week? We were in New York for the FIGMA ipo. It was a lot of fun.

Slow week this summer.

Yeah, it's been a lot. People are always.

Look, people always say it's like, you know, it's like, oh, it's really busy. I'm waiting for it to slow, but it never actually slows.

It never slows. Well, speaking of things that are. Speaking of things that aren't slow, I want to talk about fast cars. Everyone at OpenAI just received, allegedly, the rumor is a $1.5 million bonus for being on the team for more than two years. I want you to help me advise these OpenAI researchers. We want to. We told them, don't buy a house, buy a car. And so I pulled up some options that you could pick up for $1.5 million.

I think what they really should do. Everybody. Everybody should get some leverage. Leverage. Take the 15.

Buy an F30. Buy an F40. Okay, so we're skipping the SP3. We're skipping the Valkyrie. We're going straight to F40. That's great. I love that, by the way.

There's at least somebody on that engineering team that will do it. Most of them probably buy, like, Priuses, but there might be, like, one or two guys that'll do it.

I mean, when the boss is driving

a Koenigsegg, you know, you got to show up in something, at least a depreciated Veyron.

Well, the thing, it's funny you said about Koenigsegg. He can drive it as long as it turns on and runs.

Yeah. Any advice for the Koenigsegg owners in the audience keeping that thing running?

I like the.

I mean, he also has F1. Well, nothing wrong there.

Right. Okay. As much as I want to talk about this. I do, I do. We have limited time, and I do want this week.

By the way, you need to have Zach Brown and I on together sometime. I'll talk to investing. He can talk to cars.

Okay, fantastic. We'll do that.

So reaction from the timeline this week is that it feels like model frontier models are plateauing a little bit. And I think that's generally fine. There's still a lot of capabilities to unlock. But I get concerned for the VCs that have been deploying billions of dollars into the longer tail of companies that have been kind of like more. They don't have necessarily traction, they don't necessarily have truly top talent, and they were kind of moonshot esque bets on this idea of super intelligence. And so anyways, I'm curious, how cooked is the venture industry and VCs broadly? If. And the other factor here is, like, within a bunch of subcategories, there's five or six heavily funded Players going after the same opportunity. And that will create healthy competition, sometimes unhealthy. And I'm sure there'll be some good outcomes. But how are you thinking about the current state of venture?

Yeah, that's a great question. So I was just writing some notes. Well, obviously a bunch of questions in there. So AI saved the venture industry from a big awakening that was about to happen in 2022 and they just got like one more kick. What's absolutely incredible is that none of these people like remember, none of these people remembered what happened in 2021. Okay, maybe you can excuse some of them in 20 and 21 because they didn't remember what happened in 99 and 2000 because like a lot of them weren't doing it back then, they weren't in the industry. But most people that are doing this today were in the industry in 20 and 21. It is the exact same thing. As you obviously know, we focus way more on profitable businesses. Like 70 plus percent of our companies are profitable, some are in Silicon Valley, but we find companies in Ames, Iowa and Niswa, Minnesota. We're probably the only tech fund on the planet that has two investments in Sarasota, Florida. Go figure.

You guys dominate in Sarasota.

Yeah, dominate in Sarasota.

Yeah. But I felt that over the last year where.

Let me continue. So 90% is okay. Before that. People always underestimate technological change in the long term and they always overestimate it in the near term. Think like the Internet bubble, think mobile phones, think self driving cars, PC revolution. Like it always, they always do that. This what we're seeing in AI is so reminiscent. And what you're seeing in the markets right now is so reminiscent of 99 in 2000. Yes. By the way, the big hyperscalers have giant amounts of profits and like are printing money. These AI companies, a huge amount of them, you know, 90 plus 90 to 95% of these AI application companies are going to zero. And by the way, title and I'm saying that I'm hearing that from some of the world's foremost early stage venture capitalists that are telling me at the end of the day 90 to 95% of these things are going boss. And a lot of it is driven by upside down unit economics. It's funny, you see some of these AI companies that have, I'm not going to give names, that have only 50 to 100 employees but are raising like $500 million and you can't spend that much money. Obviously it's because they have massive Negative gross margins. And that money is flowing through the AI companies, which is flowing through the hyperscalers and flowing through to Nvidia. People ask us what are our thoughts on the models. We've always thought for right or wrong, that the models will commoditize and it will become a game of who's got the best infrastructure and who can deliver the searches the cheapest for the life of us. I can't figure out why Google, Amazon, Microsoft and Apple, which hasn't played yet, but I think they're going to go acquire. I think they should go acquire somebody. And why don't they just win in Facebook? Why don't they win at the end of the day, I mean, Nat Friedman is a total stud who's running AI. I mean, we backed his first company, Xamarin or his second company, Xamarin. Total stud. These companies can spend 680 to $100 billion a year on capex while having margins go up. Still growing 22% a year, printing money. Like I don't know how at the end of the day, Anthropic and OpenAI and Perplexity can compete against this. I mean, I joke that Google should literally run searches on ChatGPT and bankrupt them. Like, I mean the cost.

Well, I mean the news, the news that came out in the last 24 hours is that Google is making Gemini free for students for the entirety of the next school year.

So they will be subsidizing.

By the way, it should make it free for everybody. And by the way, I read somewhere the cost to serve the exact same search on Gemini versus ChatGPT is some crazy percentage. It's like 1/10 the price or 1/20 the price. By the way, that shouldn't be that shocking given that over the last 20 years Google has spent a lot of their efforts on taking their infrastructure and making a fraction of a fraction of a penny more efficient. I think this becomes an infrastructure play now. I think there's going to be insanely amazing companies using AI that are built. I actually think a lot of incumbents are going to win. Ever since this device came out, there's only been four companies built that came out after this device that have been worth 100 billion or more. It's like ByteDance and Pindodo, two companies in China and ByteDance maybe one of the best AI companies on the planet. That is like the hidden giant out there. And then airbnb and Uber, that's it. The incumbents are going to win Salesforce.

OpenAI.

OpenAI is not the Incumbent like it's one of the new guys. I challenge the people that ask why is OpenAI not like Excite, Lycos or AltaVista? I'm not saying it is, but I think it's hard to make a bet at a 300 to $500 billion valuation that I think you either make a 0 or a 5X. I think it's really binary and the unit economics on these companies is massively upside down. If it wasn't you would see some of these companies go public. And I actually joke that people, you know, we might as well be in a bubble. In the Internet bubble nobody gave a damn that the unit economics were upside down. Take them public, let's see if they.

Yeah, I mean they weren't, they weren't.

The unit economics of Google were very, very good going into that IPO and they maintained a good unit economics.

Well, I think our general read was, was you know, journey that model like intelligence is plateauing but it's not necessarily bearish for OpenAI because it's a habit. They have 100 million people in the US that are using the product every sing and I think what you should

adjust though is that's the upside. The downside is 100 million people use it but the downside is that investing here you're going to get massive amounts of dilution because you have companies like Facebook and Microsoft and Amazon and Google that now pay top engineers like they play in the NBA and can fund in these companies you're going to get massive amounts of stock based computers. But I use ChatGPT all the time. Yeah, but I think people might, it

has Google level adoption but, but the economics are not, are nowhere.

The economics are not there. And what's crazy is like people are like well you know Lycos and like Excite and altavista didn't have that many users. I'm like yeah, but the number of users on the Internet went from like you know, 200 million to like 5 billion or something. So you need to adjust for that too. But no, I mean the consumer growth rate is insane. If you have insane in consumer growth you should be able to attract world class talent.

Yeah, but if you hear about people

don't the competitive dynamic like what OpenAI as a company has been able to accomplish despite competition from all the hyperscalers, despite competition, the talent wars, all these different things in competing in the most capex intensive industry since the railroads or

might be more capital inventive.

Yeah,

it's truly incredible. And to do that as a company that still some people see as like a startup, you know, is amazing.

I think the other. What about, I think what's interesting is you hear the application companies all say, well, the costs are going to plummet for the cost for AI stuff. But the issue is if the costs are going to plummet, that would be good for the application companies. But then if the cost plummet isn't that bad for OpenAI or Anthropic because their revenues are now going to, well, Jevons paradox.

If the costs go down, people will

just, I mean, right now OpenAI has tons of paying customers and people that use the product. So costs like, like they are both a victim of commoditization at the model layer and a beneficiary of commoditization at the model layer. And so like, yes, you should maybe discount their, their model API business more, but then you give more value to their consumer business because it has higher margins, better unit economics. It starts looking more like Google and less like us, I guess.

Yeah.

What about, what about legacy SaaS? People have been extremely bearish. This week's in particular, it's been kind of.

By the way, that's why you go invest in it.

Yeah.

Steve Cohen has told me for years, like, if everybody's going one way, go the other way and you'll make a ton of money over the years. I mean, I think like the Chinese small cap Internet, Chinese small cap index is up like 50% or something this year. The great thing is like I don't have to make an investment in these AI, in these like model companies. By the way, I should have hit 30 billion like we should. We were wrong. You could have sold now. Right. So I don't have to make an investment, but I do run an investment firm, so I need to invest in things. And look, we look at a lot of these AI application companies and they grow crazy fast. But we're not going to pay those valuations and I'm not investing in businesses that have 50 to 60% gross dollar retention. Like screw nets.

Look at gross.

Like that's what really matters. And so what are we going to go do? Let's go find bootstrap software companies that are in College Station, Texas and Toronto, Canada and Sarasota, Florida. Companies like Pacemate that make cardiac monitoring software, companies like Gravity that make budget planning software, and companies like a list of plan that make financial planning software. Again, these companies aren't going to change the world, but again we can take them from 20 million of revenue to 60 million of revenue and There's a lot of homes inside strategics or private equity funds. And by the way, these are businesses that have like 95 plus percent gross dollar retention that have been bootstrapped. Businesses that we're now using AI to make them more productive. Like they have 15 software engineers now they can use AI to have 30 software engineers. And as long as those companies still sell to humans on the other end of the table then if companies, if AI companies are selling to other companies that use AI and they're talking to agents and we don't need people, I think that's possible. But then I think a lot of these companies are in trouble. But we think a huge amount of these software companies that are out there that exist will actually use AI to create new products and become way more efficient and more productive. And I truly believe AI is going to be as big or bigger of a productivity boom over the next 20 years. And that like than the Internet was.

Is that, is it more of like a consolidating force or a decentralizing force? Is or should we see more roll ups going forward with a bunch of these?

I think it also is like we really like vertical application software. Like all those companies I mentioned were very verticalized. They're like own a system of record and like they can just use AI to become like I know some, there's some very smart private equity guys that are using AI to try to start to do roll ups and stuff like that or like I know that I think the general catalyst guys, the Thrive guys are you know, looking to buy insurance companies or.

Yeah, I had, I had, I had a meeting or dinner with, with a few guys. I think it was Tuesday night that we're doing AI enabled roll ups in like a couple key industries. And I won't say them because I don't want to give away their alpha. And then Wednesday I had a meeting, John and I had a meeting in the morning with somebody that was like yeah, like you know, at least one of the labs will hit escape velocity and we'll get a fast takeoff in the next two years and nothing's going to matter. And I was thinking to myself like even in that fast takeoff scenario, like it's just, it's impossible to imagine I'm

still going to be using a couple

of these categories, you know, a number of different categories where like people will just maintain control, right?

Yep.

As long as. So we have, as you guys know, our LP based all these like world class execs and entrepreneurs and people like that. And we talk to These people that have built giant software companies. And one of them said, I was like, do I, do we need to be worried about all these companies going bust? It's like, well, as long as your software company sells to another human on the other side of the table, like, that human will need to like, make an interaction and probably interacts with software so you're okay. And he's like, well, unless you think like an AI agent is selling to another AI agent, like, if Lead Edge doesn't have any employees and my AI agent is talking to your AI agent, then maybe we don't need people. But as long as people are involved in the equation, you're going to need software companies.

Yep. Makes sense. Well, we are running behind. This is fantastic. We got to have you back on the show anytime. Thank you so much for having me.

We'll go for a drive for sure

and come back on. I want to get a sense of how we should be thinking about IPOs in this back half of the year. Obviously, Figma was very exciting. There's more on the horizon and I want to get a sense for how

you're kind of happy to talk about it. Funny thing is, if you go back and look at the last, like, everybody said the IPO market was dead. It actually hadn't been dead. It was dead investors. But you can go look, look at Reddit. Look at all these things that went public over the last year.

They did great. Yeah. The IPO window secretly been open for. It was open.

Admit that they overpaid for the company, didn't want to take it public.

Yeah.

You have companies like, we're in this. We're very early investors in Grafana, but you have companies like Stripe and a bunch of things that are awesome businesses.

Yep.

Peter Bricks. They just have so much capital on the balance sheet. They don't need to go public.

They don't need to go public. And so. Yeah, yeah, yeah, yeah. And so there's very different than the window being open versus the time is right for the best companies to actually go out. And they make that short because it's nice. It's nice to be in the private markets.

You mean it's not fun for most funds to not have to mark their books every day?

Like, yeah, that's fun. SEC filings are not fun. Tender offers are fun. There's plenty of ways to stay private and have fun.

And actually what's amazing is, is like we are finding opportunities. Crazy. Where you can find opportunities. You'd be like, how does the same size company in Sarasota, Florida. Get done at five times revenues. And a company in Silicon Valley, but it's growing faster. It gets done at 90 times revenues. The world is becoming this like there is true value to be had by just looking a little bit off the beaten path in our view.

Yeah, yeah, for sure. Anyway, thank you so much for taking time. We will talk to you soon. Have a great day. Have a good weekend.

Have a great Friday.

Bye. And next up we'll bring in Ben from Orbital Operations. We gotta get the gong ready.

Get it ready.

Ben, how you doing? Are you there? Sorry to keep you waiting. He's going great to have you on the show. We got some folks in the chat. The IPO window has been open, but everyone hasn't given the group of new public companies credit. Everyone is treated as a one off. Like it's not indicative of collective market. Good point. Dan Ratliff. Anyway, welcome to the stream. Do we have Ben? How you doing?

Yeah, sorry.

I appreciate you having me on the show. Yeah, I was listening to the previous conversation, it was pretty interesting.

Fantastic. Thank you for hopping on. Would you mind introducing yourself the company and then what news you got for us?

Yeah, so my name is Ben Shlonager. I am the co founder and CEO of Orbital Operations. And we just announced closing of our seed round. We raised 8.8 million.

Congrats.

There we go.

Thank you.

Thank you.

Who's in the deal? Who's making money off of this?

So initialized capital is leading around. We've got a big participation as well from Harpoon Ventures, DTX Ventures. Rebel Fund is in as well and a lot of other angels that joined on board. Yeah, we went through Y Combinator earlier this year.

So Y 8.8 was there.

Good numerology. Chinese number wealth.

More.

More of a.

More of a range we were trying to hit. Based on the technical milestone, Orbital Operations is developing a high thrust space vehicle. It'll be stationed out in orbit for satellite defense. So our, you know, we've got some, some hard tech development to go through and really prove out our technologies.

So satellite defense. I have a satellite up in. I have the Hubble telescope. I don't want someone to shoot down the Hubble telescope. Telescope. I pay you to loiter around and blow up any missiles that are coming at it. Like what? How are we actually defending?

Yeah, yeah.

So it's interesting. We have a ton of critical. And by we, I mean the United States has a ton of critical infrastructure out in, you know, both low Earth orbit, but even higher out like medium Earth Orbit, geosynchronous orbit. You could think like gps, naval communications, nuclear command and control, they all sit at these higher orbits. It's actually really, really challenging for, like, missiles or anything to get out there. Rockets designed to go out there.

Yeah.

But there are adversaries placing satellites that have capability, and it's been demonstrated already to be able to grab other satellites and pull them out of orbit or, you know, be able to fly a solar panel or jam communications, whatever it is. And we currently don't have a response for this. We don't have anything stationed out in these higher orbits. And so that's what we're really looking to build, is a vehicle that has enough thrust, fast, enough response time, and enough extended range to be able to go and intercept these things, hopefully not blow anything up. Space defense is a little weird. You don't want to create shrapnel. Right. The last thing you want to do is killer. Killer shrapnel. Yep, exactly.

Walk through. If you don't want to create shrapnel, you don't just want to run into this. I'm trying to comp it to Android. You know, you have the anvil. It just like it's a stone, that rock that runs into the drone. Then there's, you know, remote takeover, the microwave. Different radiation, different. Different energy sources. Jordi, we've had a company on a. On the show that just. Is a gun on a truck. It just shoots down the. Creates a lot of shrapnel when it shoots down the drones. But there's, you know, I've seen, like, eagles come and pick up drones. Walk me through the different tools in the tool chest for taking out a satellite killer.

Yeah, yeah. So, I mean, unfortunately, one is to just, like, ram into it, and that is. That is kind of the worst case last resort. You make a bunch of shrapnel. Here on Earth, shrapnel falls to the ground easy, but up there, it keeps. It stays there, runs into other satellites, creates a problem. The next is kind of what you were mentioning, you know, direct energy type of stuff. This would be high powered microwave. This would be frying a solar panel or a laser into the cameras and sensors, trying to basically degrade the satellite.

Right.

The other option is to do what we would call an rpo, or remote proximity operation, and actually go up next to it and grab it. Right. This is a more challenging operation to do, not something always kind of considered on the defense side, but. And more on the logistics side of things. But it is something that you could do as well.

So grab it, grab it.

Take it out into deep space.

Yeah.

Fly it into the sun.

Yeah, I'm sure that'll be easy.

Well, I guess my immediate question is, I mean, it feels like incredibly important work. And also how do you kind of like prove out your team's kind of like capabilities? There's not exactly like a test range that you can take. You know, we'll have defense tech founders out here where they can just go out and do a demo for the army or the navy or whatever.

Got to go to the moon.

And in this case, you know, how do you, how do you kind of test out different theses at the product level to prove capabilities when you're not in a, when, when we're not in an active, you know, conflict.

Yeah, yeah. I mean, it's similar. You know, my background is all rocket engine development.

Right.

So a lot of it does start on the ground. Yeah, well, we'll go rocket engineer, rocket scientist. You know, it's, it's, it's, it's too flattering. I can't, I can't take it.

We need rocket researchers. We got to get these salaries.

There you go.

Rocket researchers.

That's got to hire a bunch member,

member of rocket staff in the bio anyway.

But yeah, you develop it on the ground, you're hot firing your rocket engine here on the ground. You're developing the tank, you're going into thermal vacuum chambers and testing all of that. And one of the great things about like this day and age is it is actually getting cheaper and cheaper to get into orbit and start demonstrating these things, you know, ride shares. Maybe you do a subscale version of what you're going to do, which is what we're planning on doing for our orbital demo. And you do a ride share and it is cheaper and it is faster than it's ever been. So it is not quite as easy as just going out to a live fire range. But at the same time proving out your core piece of tech, it is getting easier to go to low earth orbit and test it there.

We're running behind today. So last question from my side. Is cost to launch actually still dropping on some sort of exponential scale or are we in a plateau? Are we waiting on starship to get through testing? Like, are we actually reaping the benefits of cheaper and cheaper launch costs or are we kind of in like a local plateau?

Interesting. Yeah, I mean, I think we've been maybe in a little local plateau for the last couple of years, But I mean, SpaceX has been dominating the market. Right. I think there are other launchers that are coming to market that'll help drive competition down. And then you have Starship and Stoke bringing reusable second stages. That will be another step down. That actually doesn't make it cheaper getting out to the higher orbits. It makes it cheaper getting to low earth orbit. But those higher orbits, you know, the reusable rockets and getting out to those higher orbits, they're not good at doing that necessarily.

And that's where an impulse space comes in. Tom Mueller's company, that is a booster that takes you from LEO to. To. Yeah, Kickstarter Kickstage. Got it. Very cool. Jordy, do you have anything else? Because we're sorry to cut this short, but we are.

No, no, no.

Totally fine on a Friday. But thank you so much for doing jumping on. This is really fun.

I came to gong, so I appreciate it.

We hit the gong, we'll hit it again for you. Let's go.

Yeah, appreciate it. There we go.

I gotta get one for the office. You need one. Well, good luck. Come back on when you put something in space. We'll ring the gong again. We'll talk to you soon.

Great. Thank you so much.

Bye. And up next, we have Meryl from graphite coming in. You know graphite? We do. We read ads about graphite every single day.

And we use graphite while building my last startup.

While he's joining, let's tell you about a different ad sponsor. Bezel. Go to getbezel.com your Bezel Concierge is available now to source you any watch on the planet. Seriously, any watch get a hitter. AGI is delayed. Super intelligence. Whether. When it. When it gets here, if it's five years, if it's 10 years, it's going to want to look on your wrist and know that you mean business.

That's right.

So go to Bezel.

Well, without further ado, let's put bring in Meryl on the show.

Meryl, how you doing?

Doing well. Thanks for having me, guys.

Good to catch up with you. It's been a week. I saw you in New York City, had a lot of fun chatting. Give us the update and give us the actual impact of GPT5, the news that happened yesterday. I want to kind of noodle through that and the impact what it means for your business.

Awesome. So GPT5 obviously was a huge announcement, made a lot of waves. Our team immediately got to work on testing it and I think we're noticing a few things that are improved about it and a few downsides. That come along with it as well. On the improvements side, I think it's a lot better at deep thinking. It's good at one shotting apps. It's also the biggest surprise for us is it's meaningfully cheaper on inference than a lot of the previous models. On the downsides though, I think there's a lot of buzz around it being this massive step function. And for us, what we've seen practically in reviewing code is that it's an incremental improvement, but it's not this massive step function. It is very much still in the realm of many of the other state of the art models. The other piece there that a lot of folks have commented on is latency. I'm sure the OpenAI team is working on improving this. But one thing that we've noticed in updating our code review product.

Yeah, I want to talk about latency

versus using just like the previous generation of models that are now being deprecated.

Yeah, the previous generation of OpenAI models, anthropic and others.

Yeah, I want to talk about mixture of models. We were in the mixture of Experts era with GPT4 and then Grok is doing more mixture of different models together. And I'm wondering, like code review is pretty high stakes, it's pretty high value. It's a very economic like, like, I don't know, like the average salary or hourly rate of someone who's capable of doing code review as a human is really, really high. And so is there a world where I actually want to run within graphite code review on every single model and then have you design a rubric or a scoring system that compares Claude to GPT5 to graph to all the different models and kind of lets them war it out and sits at a higher level of abstraction. Obviously that's more expensive on the inference cost side. But is it cost prohibitive compared to a human reviewer that might be a couple hundred dollars an hour?

Yeah, it's a great question. I think today it's certainly cost prohibitive to run every single model. What we do though, and what's worked really well for us is we much like a human reviewer, we break down the task of code review and we're looking for different set of things at different times. So we'll look for bugs, security vulnerabilities, efficiency gains that we can make code based style guide inconsistencies with the rest of the code base. We also let customers define. Every team has their own protocols and their own guidelines around how code should look and we let them define Custom rules. Many teams really have like heavily leveraging this now. And for each of those tasks, their own tasks in the review process, we also then once we generate a lot of comments, we'll have simple things that we're looking for, like little rules. Even these can be things that we've learned over time, like if left to their own devices, the models like to say things like we should update this line of code or something. And developers find that really annoying. So you probably don't need the highest power model to say, you know, we should don't add comments that start with we. But over that is how we've kind of composed the logic and composed the, you know, the voting system that we use to determine is this actually a comment that's going to add value or is it just going to be noisy like many. The challenge of many products and especially AI code reviewers out there.

Yeah. How do you think about decomposition? We've heard this trend of like the. The smartest frontier models might be training smaller models. We've talked to a couple like almost like micro foundation model companies where they're training, you know, it's just a great model for filtering for profanity and it runs and it was trained on like video game graphics cards. And I could imagine in the world of like tool use there becomes a future where there's a bunch of small models that are being kind of orchestrated by the more expensive model. How do you see kind of the surface area of what you're building kind of fork out?

Yeah, I think that there's especially as we think about the scope of code review, it is kind of this first collaborative moment in the developer lifecycle. Historically we've seen co generation be fragmented, every developer having their own IDE and terminal setup. Even today we see this with a lot of our customers that are using. Some engineers are using cloud code, some prefer cursor. Everyone's trying the new cursor CLI now there's so much heterogeneity on the code generation side, but that's always been fragmented. But code review has always been unified and it has to be because everyone is working together. It is that first collaborative moment in the developer process and then it connects to all these other pieces around CI merging, deployment, deployments, everything that comes downstream of that. The moment you create that PR to get it out to production. And each of those I think represents the different tasks to be done to move those along. So you could think about having one model or one agent that's really good at resolving merge conflicts and Another that's really good at looking at CI failures and figuring out what the problem was and being able to just fix that in the background without you having to do anything. And we see that. I think the impact of that will mostly be one of reducing costs over time and just letting you run a smaller model and not have to use this super powered laser on even a little task that something smaller could do.

What does market share look like from your view upstream of graphite? So yesterday we had a bunch of people on that were that had gotten early access to GPT5 and everybody's heavily conflicted because a lot of them are doing code generation in some way or another. It's hard to really suss out like okay, what is actually dominating other than just picking up what people are saying on the timeline as actual users. But what are you guys seeing?

Yeah, we're seeing a pretty big shift in the past. Even over the past three months I think we've seen this shift from primarily orgs using like using cursor and moving more over to cursor to Claude code I think has really started to dominate the conversation for companies of all scales for the past few months. And now we're seeing, just in the past day or so we've heard a lot more interest in the CRISPR cli. I think that Claude really proved out that prompt first modality working really well. And we've seen this sort of shift from the code first modality of tools like Copilot and then Cursor and Windsurf to now prompt first with Claude code Cursor CLI and then I think the next big shift is staying prompt first but moving from local to remotely deployed agents. And we're already seeing, we've seen cursor build something there, cognition, codegen, many others that play in that space. So I don't think that we've seen a massive, I don't think we've really seen a massive shift in the past 24 hours though. It's still, it's still. I think the good news for everyone is that it's still a pretty close race and competition is really I think necessary in this world. I think if we ended up in a place where there was one model provider that was so much better at cogeneration than another one, then there wouldn't be as much pressure to innovate. They'd have a ton of pricing power. It really wouldn't. And it wouldn't reflect that historical model of like co generation being kind of up to developer preference.

How is AI changing developer communication around pull requests. Like the canonical example is like you write a summary, you write like a headline. But I imagine that it's probably pretty easy. Like one of the things the models are great at is just, just condensing down information. I'm not a particular fan of them expanding information oftentimes, but, but, but if there's, if there's a really big poll request, you can kind of show varying levels of summaries. And I find myself doing this even in the consumer realm where I will say, okay, I want a deep research report on something, some, some the history of a business. But then give me a one line summary, give me, you know, five bullet points. Then give me, you know, a New York Times article length, a couple hundred words, words and then give me the full 30 page PDF and because I want to be able to consume it like a fifth grader, like a college student in these successive levels of depth. Is any of that happening in the pull request world? In the developer code review world?

Absolutely.

I think one of the most used AI features and one of the first things that we launched in the graphite platform was the ability to write the PR description for you. And it's something his developers kind of famously hate, writing long description. Yeah, yeah, just do this, do this for me. But AI is amazing at understanding a change. Summarizing it. Now we're actually, I think this is also where we'll see code review going is it's moving from this world where you're just scrolling through the diffs and everything is just an alphabetical order and you have to kind of guide your own way through it to. I think now AI is incredible at understanding the change. Like we were actually working on, working on a new feature launching pretty soon in graphite where you'll be able to just ask graphite, hey, what are the important parts of this change? Like walk me through the key pieces of code that changed here. What should I be looking at? What is high risk and making it a lot more. I think even in AI code review right now, I think we're still in the Copilot v1 moment of AI code review where we're just adding comments on GitHub. But the future, I think is much more one that's interactive and is guiding you through the code change in real time and helping you to both review it and also to make updates and coordinate the various agents that are working on that change in real time.

That's great, Jordy. Anything else? I think we're.

Yeah, I'm curious I mean the main thing is, I guess like going into this year, everyone said this is the year for agents. And I was. And it felt like yesterday with Greg or Mark, I said this was the year more of deep research. But then they said maybe also coding agents. Coding agents have been the other thing where what do you expect out of the coding agent market in the next, like before the end of the year?

Good question. I think that the biggest thing is that we're just, we're just moving through, like slowly moving, moving around that grid. If you have like the way I think about the market is you have one axis of where does the code live. On the one hand, historically it's all lived locally. Now we're seeing a shift to some of this living remotely and having these background agents that you access primarily through prompting. And then on the other axis you have is the interaction modality code first or is it prompt first? The code first ones being you started with copilot, you had cursor Windsurf. Now we're seeing the prompt first modalities and Claude code, the cursor cli warp and others. And I think we're starting to see, we've seen that shift from the local code first to the local prompt first tools now and now I think what I'm curious to see for the rest of the year is how quickly we then go from local prompt first to remote prompt first using cognition, codegen, cursor, background agents, those others. And I think our bet and what we're seeing is that that is very much the future. It likely requires another step function improvement in the models for that to truly be the primary modality of software development. But I think that will be the shift that we'll start to see for the rest of the year is that migration from local to remote prompt first

modes makes a lot of sense. Well, thank you for joining.

Thank you so much. Have a great Friday. Have a great weekend. We will talk to you soon.

I'm sure it'll be a busy weekend for the whole graphite team getting making sure that GPT5 is rolling out.

Fantastic. We'll talk to you soon. Have a good one.

Cheers. Thanks for having me guys.

Later we have to talk about the actual biggest news in artificial intelligence. We missed this. Everyone's been focused on GPT5. The biggest news in artificial intelligence is that the billionaire AI co founder Lucy Guo pays nearly $30 million for an LA spec house. Lucy Guo co founded Scale, now runs passes got a discount on a Hollywood Hills home tech entrepreneur Lucy Guo 30 has been called the world's youngest self made woman billionaire. Now she's putting some of her earnings into real estate. She paid $29.5 million for a newly constructed mansion in LA's Hollywood Hills. According to people familiar with the transaction, the price is a significant discount. They originally were pricing this house at $43 million. It's a 13,500 square foot house, five bedrooms on 1.2 acres. And this is, I mean everything about this house looks great. The photos are fantastic. The one mistake. Lucy Guo couldn't be reached for comment. Why are you no commenting this article. You gotta tell your whole story.

All I can say is you can sign up@passes.com exactly.

There was, there was a great, there was a great opportunity to upsell to do all sorts of, of stuff but there was a I. This, this house has everything. A sunken, a sunken fire pit, a pool, jacuzzi. It has everything. It looks fantastic. LA's high end real estate market has been hobbled over the past year by the LA fires, outward migration and lower to lower taxes and broader economic uncertainty. Most of the recent high end sales have closed at significant discounts. So there's really no reason to buy a 30 friend reason not to buy a 30.

Friend of mine in Malibu says that there's a barbell in the market right now where things are hot under, under like five like single family homes and then really hot on the high end in like 50 plus range.

Well, speaking of 50 plus, ex Google CEO Eric Schmidt has purchased LA's Spelling Manor for 110 million million dollars.

Finally.

This is a very, this is a very fun article. Amids a challenged luxury market, the property sold for less than its 2019 sales price. He got a steal. This was built the Spelling Manor. If you're not familiar with this, this is an iconic legendary house. If we can pull up the image. Look at this thing. It is insanely large. Multiple wings.

It has an entire like a B21 writer from above.

It's amazing. It's amazing. It hasn't, it has an entire. Famously has an entire room for just wrapping presents. Because when you're at this level. Yeah when you're at this tier you don't just like you got to be wrapping. Yeah, you need a whole room for wrapping presents because you'll be wrapping lots of presents because you'll be giving out lots of presents. It's a 56,500 square foot property, 14 bedrooms on five acres and this is in Los Angeles. This was built by Aaron Spelling who was a TV producer in the 90s. Very, very famous. The French Chateau style property is slightly larger than the White House. It has a bowling alley, a wine cellar, and a beauty salon with massage and tanning rooms. So if you need to get a tan on, you go to the tanning room. I don't know what the movie theater status is. We need to see. While the property will remain a single family home, the Schmidt's philanthropists, who have homes around the world, purchased it primarily to host meetings and events for LA nonprofits and cultural institutions. So they're not even. They're not even living in it full time. It was sold for a discount. It was listed for 137,000,000.5 after several years on the market and multiple price cuts. I mean, it's such a big, iconic house, the buyer pool is probably pretty small. It was last sold in 2019 for about 120 million when the seller was British heiress Petra Ecclestone. Ecclestone famously hired a team of roughly 500 workers to complete a massive renovation of the property. And now, just six years later, the Schmidts are planning a significant remodel of the house to simplify the floor plan. So this place, this thing is just getting more complex. And then they're putting it in a present wrapping room. They're tearing out a present wrapping room. Everyone is remodeling this property. And they're no longer going by Spelling Manor. You know, the guy who built it, he put his name on it. They're pulling it off, and they're calling it 945-94594, a reference to its address on Mapleton Drive.

Very nice. Well, I have a post to cap off the week. It is in the timeline from Lulu Maservi. So we've been, I guess, mentioning over the last few weeks that it feels like we've maybe reached a local top on startup launch videos. A lot of them are starting to look the same. They're sort of the default now. It's very hard to stand out. People, I think, are starting to glaze over them a little bit. But Lulu has potentially a new meta. So let's pull this up. She says, who's gonna be the first startup to do this format as a launch video? And it's the ferocity. And I would go out and bet that it's Clulee, of course, that they're

already working on it.

They're probably working on it. There's not a full map that they don't like.

The challenge is some startup might. Might.

This is so much choreography. I guess the people in the. In the back are just kind of dancing up and down. They're not really.

Yeah. So the challenge is if you're launching your startup Monday.

Yep.

You could be spending all weekend doing this, getting it ready, and then Roy Lee and the team will launch it, like, Sunday, probably.

Probably.

So good luck out there if you're trying to.

You know, speaking of Cluli, obviously they still have to do so much on the product side to really, you know, nail it, get retention, all this other stuff. But in terms of just the cheating as a keyword, I think that there's something that's going to be very, very sticky there. Julia Steinberg has been on the show, is posting a picture of a billboard, says, hi, my name is Roy. I got kicked out of school for cheating. Buy my cheating tool cluli.com and I just feel like it's such a simple encapsulation of the value proposal just in one word that grabs your attention. I was thinking it's kind of like Red Bull gives you wings. It's not Red Bull doesn't actually give you wings. It's just.

Speak for yourself, John.

Yeah, yeah, you've grown wings. You've grown literal wings, you know, but. But it sticks in your mind. It grabs you. And I feel like the. Even though the cluely stunts are kind of getting, like, less and less, it's less shocking because it's like, oh, okay, we kind of come to expect Cluley is going to do a sorority dancing video, for example. But just the cheating as a keyword, if they can own that in the broad consciousness, I think it's going to continue to pay dividends and grab people's attention and just be something that they can run with for a long time. Because even though the Cluli stuff went super viral, everyone in tech talked about. Everyone in tech was arguing about it for a long time. There's still hundreds of millions of Americans that have never heard of cluli or maybe they saw it once and forgot about it and just. And I think that that is going to be like a keyword marketing campaign that just, like, keeps sticking in people's mind, and it will find out saturated

facts in the timeline, and they have to figure out how to go mainstream.

But, like, it's such a distillation, it's such a compressor impression of a meme of an idea of, like, oh, I would like to, like, cheat on that, you know, report that I've been working on at work. And actually, my boss doesn't care if I cheat. They don't care if I use that. They care about the result. And so. But it grabs your attention. And so like I could see it. I could see it running in a Super bowl ad and actually driving conversions and downloads. Now the product has to be really great. It has to be better than chat GPT. That's an extremely tall order. They're going up against Gemini, which will help you cheat, quote unquote. I'm using cheat in the just AI assistant language. Not actually literally cheat, but. And they're going up against free models and cutting edge models and all sorts of different things. And we'll see how sticky that new their UI implementation is where it screen scrapes and records your calls. But in general, I think just as a marketing technique, I don't think we've seen the end of cheating as a buzzword that clearly will be reaping the

benefit of totally any code.

Anything else in the timeline, Tyler? Any. Is the timeline in turmoil? Is the timeline quiet front?

There's one thing. So some small updates for GPT5 from. From Sam Altman. He says they're doubling rate limits for plus users.

Can you do the same? Altman voice?

What is this?

They're doubling. We're doubling rate limits.

We're gonna let PLUS users continue to use four.

Oh, they're. There we go. That's good. That's how I imagine.

Yeah, they're bringing four zero back.

Interesting.

And the auto switcher was broken yesterday.

Oh, interesting.

So he says it will seem smarter today. The result, it was out of commission for a chunk of the day and it will.

It's kind of hilarious that they had like, you know, pretty big launch. There's like two bugs. Like one with the chart clearly got mis rendered for the live stream was fine on the blog post. And everyone's like, chart crimes.

This is.

And then the model switches back broken and people are like, this is. Doesn't work at all.

But I'm thinking, I think we're back.

You think we're back? I think we're back, baby. We heard it.

We've heard it.

You heard it here first. We are back. Accelerate your timelines. You now have two days to escape the permanent underclass. So have a great weekend, everyone. Get to work. Two full days. I. I will be. You know, I think, I think most people. It's summer. You should escape the permanent underclass this weekend by hitting the pool, hit the beach, have some fun out there.

Get a ramp floaty.

Get a ramp floaty and spend a weekend on it. Enjoy. Have a great weekend, everyone. We will see you Monday. Leave us five stars on Apple podcasts and Spotify. And thank you for watching. See you later. Goodbye.