AI Is Running Out of Human Data. Here’s the Trade.

Dave (0:03)

Claude is now doing 26% of Anthropic's own research. That's up from under 1% back in February.

They have 30,000 AI agents all running experiments and writing code and building the next Claude. But here's the problem. Every AI lab is training on the same internet, the same threads, the same answers. And that internet is now flooded with AI slop. Models are training on models. It's a photocopy of a photocopy, and every generation gets a little blurrier. So the labs are desperate for the one thing that money can't print. Fresh, original human thought. Real people arguing, reviewing, coding, explaining why they believe something. And there's a finite amount of it. Ten days ago, a federal judge told one AI lab that it can't just scrape for free anymore. The pile of human arguments just got a legal price tag. Today on Dumb Money, AI is running out of human data. Here's the trade.

Hey there, Dave here along with Chris and Jordan. We are Dumb Money. Welcome to Dumb Money Live. Do me a quick favor if you don't mind. Smash the like button, a real human clicking on it.

As of today is still apparently valuable training data. So good morning, Chris, Jordan. We know Chris, you have been building a big position in a company that is sitting on perhaps the largest pile of human arguments on the internet. And you say you might be going bigger, is that right?

Chris (1:29)

Yeah, I mean, I wouldn't say that I have like that meaningful of a position at this moment. It's moderate, right? I mean, we all know Wall Street's paying for these machines, right? I want to know like where the human signal is that the machines cannot manufacture, right? Like, I think that's what this is all about.

You know, if you look at every piece of research in the AI space, it becomes pretty evident that we're going to get into some kind of issue here in the next few years with, you know, just AI data becoming too large of a piece of the source data, right? And you just are not going to have enough actual human data. And it just starts to build on itself. And then I've been reading all these reports, and you never know what's true, right?

Dave (2:34)

Well, that's the main problem now. You don't know what's true because we have the entire human history of information and the AI bots now have access to more and more of it, and they're processing more and more of it. But they're spitting out more information faster than humans ever created it. And it is a problem.

Chris (2:56)

Yeah.

There's a thesis out there that the data get starts to get really bad if it just becomes replicated over and over again with AI data, right? I don't even know if any of this is real. This is the problem. People are throwing out papers, opinions, thesis.

We're in unchartered territory. I don't know if anyone can say with conviction exactly how this all plays out, but it seems like at a minimum, buying or paying a lot of money to lock in one of the very few last remaining truly fresh high-signal permissioned human data is a no-brainer for all these big labs, right? How do they not do whatever it takes to acquire that? I just don't see very many scenarios where they just blow that off. Because if you look at Google, Google inherently has the data advantage, right? With video data, they have a huge advantage even with email data. If you start to look at a couple of the other labs, they don't necessarily have as much owned data. They have to go out and license it, right?

Jordan (4:16)

They don't have to. With email data though, do they have the rights to use that to train their models? Is that in the terms and conditions within the email client? I haven't looked at that one specifically. Now, XAI probably has a pretty good advantage in being able to use x.com data.

Chris (4:38)

Do you think XAI is that great of a data source though?

Jordan (4:42)

I think it's good. I was listening to a podcast where they're interviewing the CEO of Reddit and he thinks that, obviously, the best thing about their data is that it's highly human, its arguments back and forth. Sometimes when we think about AI, we think about we just want the answer.

But sometimes it's not about the answer. Sometimes it's about differences of opinions, it's about the argument, it's about different use cases, it's about the journey of trying to figure out what's the best situation for yourself. I think that you get that really good in Reddit, and you get that to a certain extent in X also, because there is a back and forth and there's a way for people to argue about different topics.