ROUGH VIBES: WHEN THE HYPE CRASHES

E92 ROUGH VIBES_ WHEN THE HYPE CRASHES

Transcript

Dalton Anderson (00:00.45) Welcome to Venture Step Podcasts where we discuss entrepreneurship, trends, and the occasional book review. For three years, we've lived in a world where OpenAI was the sun and everyone else was just in orbit. But recently, there's been a slight shift after Google's recent release of Google Pro 3.0 and some of the benchmarks that have come out.

There was also a leaked memo from OpenAI talking about, from Sam Altman talking about rough vibes and economic headwinds. And so they're feeling it. And I think that this is one of the most interesting times to be a consumer and one of the interesting times to be talented and working at some of these companies, because you don't know what's going to happen next. Each company has got to push so hard to grab that next nugget.

And I wonder what it is. I wonder what they're gonna do. So some of the stuff that they did propose looked to be very risky. So I'm curious how it all plays out. Hopefully we don't die. But I am always interested in what's gonna happen next. So in this episode, we're gonna be talking about this open AI crisis. I don't wanna say crisis, maybe a word, which is rough times. The rough times memo.

the valuation disconnect and some of their monetization issues. Then talking about Gemini 3.0, I guess the enterprise gap between anthropic Google's Gemini offering versus OpenAI. And then,

That's pretty much it. But hey, you know me, I'm the host. I'm Dalton Anderson. Welcome to the show. think we're getting into the mid 90s here. And it's so funny how in June I was super sick. And then after that I went to a product onsite for the company I work at. Then I got sick again and I was sick for a month. And then I'm no longer sick, but then I broke my toe randomly. Like I stubbed it on a brick outside, just walking around.

Dalton Anderson (02:14.302) And then after that, I was working out on Sunday, set aside my whole Sunday to work on some personal stuff, to create a list of potential guests I want on the show and reach out to them.

And I was working out, doing some pushups, some sit ups, and before that, and some pull ups. And before that, I did yoga for an hour, just to work on my mobility and such. And then about four hours later, my back got super duper tight. And I was like, that's not good. And then before you knew it, my back was locked up. I couldn't, I physically couldn't record the episode yesterday.

So I did two hours of stretching to help my back, these like special back exercises. And then today I try to do the same thing. And then I also had to struggle getting on my clothes and taking off my clothes. I think I fell asleep with my socks on because I just had a hard time getting my socks off. I didn't mean it like that, but whatever. I guess people clip it. yeah, so I couldn't take my socks off.

I also went and got a massage today. I bought some Tiger Balm and some ibuprofen and that's doing wonders. Just one thing after another. I think this is hopefully the last thing, the last thing. I'm going to be very hopeful for the last thing and lasting for awhile. And then I got a massage and that was nice. That helps out. That helps out a lot. And it was funny because it took me like 10 minutes to get my clothes back on because I'm still kind of wounded.

My back is still a little tight, but it's really loosened up. The Tiger Balm, the Massage, the Ibuprofen. I am feeling a lot better. Well, I think that's funny. I always got something on. I'm falling apart, but I'll be back. While I'm falling apart, somebody at the bottom is just putting together pieces and it's just going to be this hardened, optimized structure. That's what I'm envisioning. Okay. So,

Dalton Anderson (04:23.246) Let's dive into the episode talking about the evaluation and the reality. So open your eyes since they've been the sun, they've raised an insane amount of money at this crazy burn rate and they are valued at 500 billion. They are projected to lose 8 billion in 2025. And just on

Compute and talent, they're spending $6.7 billion.

which is a lot, right? Like, I don't know how else put it. That's a lot of money for a company that has a hard time monetizing their user base. They're saying that monetizing around, or it's estimated to monetize around four to 10%, which is low, which is part of this conversation about economic headwinds and rough vibes ahead that Sam stated. And it's just a different,

a different vibe, like the user paying for OpenAI and then the user paying for Google. I feel that they're two different things and there's a lot less...

There's a lot less tension or attention. Attention is not the right word. I'm looking for a special word. Drag, maybe there's less drag in the Google landscape than there is in OpenAI because OpenAI, when they integrate their product, they're integrating into some ecosystem where Google has an ecosystem like 200 or not 200.

Dalton Anderson (06:11.918) 2 billion users are using search and Android every day. So right then and there, if you integrate with Google search in Chrome and then you integrate with Androids, there you go. Boom. That gets you, that gets you quite a bit of users. I mean, they're not going to use it every day or it's there as an option. And so when you have an AI question,

Do I wanna download an app or do I just wanna use the app that's pre-installed?

Do I wanna go in my settings and go through a whole bunch of hoops and change my assistant to OpenEyes or do I just wanna keep it the default? Yeah, there's just so much less drag on the user and it makes it easy. mean, the thing, half the value proposition of a company is how easy it is to do business.

You could have the best product, but if it's pain in the ass, no one's going to use it. And so I'm not saying OpenAI is a pain in the ass to use, but it's harder to use than others. So people have to go out of the way. And people were willing to go out of the way when their product was substantially better than everyone else on the market. Enthropoic's always been great at coding, very, feels like, feels like the model is kind of drowning when you're speaking to it. It doesn't feel,

expressive like OpenAI or Google Gemini. And so the model just felt like it's dampened and everything is muted and it doesn't have any color or it's not very lively. It's how I describe it where OpenAI is and Gemini is. And you can kind of see that they're different user bases where Anthropic is Law and

Dalton Anderson (08:08.415) very sensitive about their model and they're all about safety and safety, which is fine. There's no wrong or right approach. It's just different routes. Anthropic is more enterprise-based. OpenAI is more expressive, lively, and can be used either way. And then Gemini seems to be more enterprise-vibe than

consumer, then they do have some products like banana that went viral. And that was like a tick tock sensation. So it's kind of hit or miss with Google, like what, what they are. think they're still figuring that out. But Anthropic is very solid. Hey, we're not going to win in the user base. We're, we're going to be good at coding and then we're going to have a very muted model that enterprise is going to feel comfortable with. And then we're going to specialize in law or something like that. And open eyes like word

We're the sun. We're good at everything. We're the best. And then Anthrobic was like, yeah, that's fine. We'll just be good at this stuff. And now Google is like, well, I want to be good at everything too. Let me play. I want to play. That's kind of where we're at. Where the memo came out a couple days after Google's release of Gemini Pro.

They are monetizing a low amount of their user base. Their customer acquisition costs, like to get a premium user is higher than a Google because Google already has all these daily users and they just got to throw something up, throw something up in their ecosystem. Well, one thing that I think is very interesting to talk about the catch up is they're actually having diminishing returns on training.

And so we're getting to the point where we're getting to the scaling wall and we're at this point, we're getting pretty close or very close to the point of maximum yield to where you have already gotten all the increasing returns on optimizing the model. And now if you spend any more money on training or time on training or make it more complex, actually it's not worth it.

Dalton Anderson (10:36.003) the model becomes overtrained or it just doesn't perform well or it's a waste of money. Kind of all those things. I'm not sure exactly how they're measuring it, but when you're training a model, you can get to a certain point where you overtrain it and it actually gets worse. And it's just like, here's the optimal yield. Like here's the curve. This is where we want to be. And so we're getting to that point where companies have trained, which is insane amount, but they've trained all the

all the data on the internet and then they're running out of data and so they're making their own data. And so that kind of puts into perspective how much, like how far they've taken things and now they're getting to the point like, hey, well, it's actually not worth it to keep going.

which I think is really cool, one. And then two, it makes things problematic for one, closed models, right? Like if you're getting to the point of diminishing return, then there should be a catch up with companies like Meta, because Meta has good talent. They've got the infrastructure, they've got the chips. So they should get to the point to where they are at the point of, hey, we're getting to mission returns here.

And hopefully if that open source model is close to what Google and Anthropic and OpenAI are producing, then why would you use a closed source model when you could use an open source model? The first thing that you should look for is open source model because the weights are open.

But I would like to see how that plays out. But the next thing that I think is interesting about this whole topic about the Domission Returns is OpenAI was positioned and is positioned as I mentioned earlier in the episode as the sun, the best, the best talent, the best models, the best everything, best ecosystem, all that stuff.

Dalton Anderson (12:43.823) And what happens when you do that is that's the expectation. The market expects you to be the best. And if you're not the best, then you're a loser. Whereas Anthropic is a position themselves as, Hey, we really like enterprise stuff and we like law and we like our safe model. So they're the safety guys. Anthropic is the safety people. OpenAI is the trailblazers, the big sun, the people with the gravitas. And

then Google is the sleeping giant that has woken up. So the positioning of Google is like, we have the data, we have the infrastructure, we've got the chips.

We were just being too conservative. fell asleep at the wheel. We'll get this right in a couple of years. And Google has made things right. Their model is consistently very competitive compared to their peers. I mean, they had to go through Bard and then they did the Gemini and Gemma and Jim. And now they're at the point to where they're at Pro and Fast. And those models like Fast is

Percosper and Compute is very, very good and it is a very solid model for the cost to use the model for both Google and if you're using a VA API as a developer. And then Perot is consistently like at the top of the benchmarks, first, second, sometimes third, but a lot of times it's first or second and it's going head to head with OpenAI. But the problem with OpenAI is they don't have the same infrastructure as

Google, it's not even close, so.

Dalton Anderson (14:31.681) Now we find ourselves in this place to where the market expectation for OpenAI was to be the best. And now that they have

Dalton Anderson (14:51.001) They potentially.

they have potentially removed the market confidence that they can execute on being the leader consistently. And so the, the footing that they sit on is shaky and it never used to be. There was no one really challenging OpenAI like to the point to where they're questioning, what do they need to do next? What's the next big question? And that's where things become very, very fun.

Dalton Anderson (15:29.847) It's...

Dalton Anderson (15:33.357) It's a place to where things are unexpected and.

you get the best product. I always talk about it, competition is so good. It's so good for customers. So if you're a consumer, it's awesome. And if you're into science, it's awesome. Because this is some sci-fi stuff. We're getting sci-fi stuff and...

It's all being pushed by capital and competition. But I mean, the capital is just half the problem. Maybe a third of the problem. The talent and the drive is really the other thing. And then I've even mentioned X and they're pushing on their own. They're trying to go to Mars. And so their model's gonna go into the Tesla Optimus. I think it's called. You little robot.

Dalton Anderson (16:25.881) So I just think it's very interesting and.

how these models are positioned. Anthropic, as I said, is positioned as enterprise. OpenAI used to lead that space. They have decreased their market share by like 25%. Anthropic has surged to become the market leader. And so there is that thought process of like, hey, let's remove the wow factor, let's just be consistent. And...

Dalton Anderson (16:59.295) If you're going to have the wow factor to keep people on board and there's other models and other companies that are servicing or potentially could service your client and they don't even have to go to a different infrastructure. They could just add it as an add on price within Google. It

It's just easier. It's just easier for...

for everybody.

Dalton Anderson (17:31.539) And Anthropic used to be the, I didn't do a good segue to the next section, but we're talking about the cloud offerings and Anthropic used to not be on Azure.

Azure used to only have

OpenAI models. Now it's adding Anthropix Cloud because Microsoft is hedging. So OpenAI has lost their exclusivity. Cloud is now the only model on AWS.

And so there's just things that are slowly, they're small things, but they're slowly stacking up on OpenAI. And then there's just massive talent war between all these companies paying insane checks, the whole thing that Meta had going to build their super AI lab. All that stuff combined with other companies creating pretty good models with less drag on the user, they have easeability being top of mind.

It makes it a very tough environment. And unfortunately, a lot of things OpenAI have problems with, they don't control. They don't control, they don't have their own chips. They don't make their chips. They don't have a substantial infrastructure. They're renting chips from Google, the TPUs, and they have their own infrastructure, yes, but they don't have their own infrastructure to support themselves outright. And so,

Dalton Anderson (19:12.707) They have that problem. They don't have their own ecosystem. Right. Google integrating AI search into their search was a big problem for them because they were going to do their own little search.

They are.

more of a integration than they are something built in. Like Google has all their stuff built in. Like it's already integrated with calendar, with emails, and all these things. And they're demoing and talking about this agentic. Like not only are you able to chat with this model, it can also handle things independently for you. And it's...

they've shown that and it's been in the plans. But to do that, it's easier when you've got an ecosystem because you've got your own stack, your control of the updates, your control of everything. If you try to do that when you're not control of the tech, it's very difficult because you have to account for all the little updates that the, that company is going to create. And then you've got to make sure your model still works. And so it's overall super difficult if you don't have that on top of

if you were gonna be in many different ecosystems, you've gotta integrate with all of them and maintain it.

Dalton Anderson (20:37.956) I still believe that they're going to pull something out, but it's definitely rough months ahead, as Sam said. And I'm not sure unless they have some massive breakthrough or they...

Dalton Anderson (20:53.924) find a different approach and maybe OpenAI takes the user approach. They just go mass users and they continue to make contracts with the government to give everybody access to AI compute. Maybe that works and then Anthropic takes the safe, muted model approach for enterprises and then Google takes the developer slash enterprise user base that's on Google and

the people using Firebase and Google cloud. Maybe that's how it plays out. But right now there's not a clear line of sight on adding substantial amount of value for people to.

Dalton Anderson (21:44.368) download other apps or change their settings, do all this other stuff that they were willing to do before because the model was that much better. But when it's close, it's like, I don't really care. It's like, do you want a burger from this place or the other place? And you close your eyes and you eat the burger, it tastes the same. So if they're that close, it's more of a personality or UI or the other things. A lot of people aren't that picky about it. If you're listening to this episode, you probably are because you're, I'm sure you're opinionated.

But a lot of people just don't care about that stuff. So they care about being easy.

It doesn't seem like they're on track for that, but thank you for listening to the episode. Appreciate you. My back was still a little hurty. I said a little hurty, hurt a little bit, but appreciate you wherever you are in this world. Good morning. Good afternoon. Good evening. Toodaloo. See you next week. Bye.