GPT-6 Astra Is Here. Are You Holding the Wrong AI Stocks?

Dave (0:02)

On Friday, OpenAI dropped their latest model, GPT-6 Astra. And once again, the benchmarks are insane. Astra handles computer tasks better than Sol. In about half the time, it scored 99.9% on OpenAI's Arc AGI benchmark. And it's the first OpenAI model to reach their critical threshold for cybersecurity capability. So naturally, AI stock should all go to the moon, right? Because for the last few years, the AI trade has been pretty simple. Better AI needs more chips, more memory, more data centers, and more power. Except Astra isn't just better. It's more efficient. And depending on what happens next, that could be incredibly bullish for some of the stocks that you own. And really bad for others. Today on Dumb Money, GPT-6 Astra is here. Are you holding the wrong AI stocks?

Hey there, Dave here along with Chris and Jordan. We are Dumb Money. Welcome to Dumb Money Live. We all know that Astra is 100% capable of just watching YouTube for you.

But you probably forgot to prompt it to smash the like button. So go ahead and do that for us this morning. Chris, Jordan, we are back on track, back to back episodes, because we need to regroup on what might just be the biggest mistake investors could make on this new GPT-6.

Being right about the technology, being wrong about the stocks. Chris, I know that you posted on X yesterday that GPT-6 Astra is, quote, simultaneously far more capable and efficient than expected, yet still far from good enough. You said that it's hard to imagine a more bullish outcome for AI infrastructure. Explain yourself.

Chris (1:42)

Yeah, I think the biggest breakthrough here with GPT-6 Astra isn't that it can do more, but it's that it can do more with less, right? So that's not actually a bearish thesis for infrastructure. That's what we've been talking about for years now. Jevons Paradox is actually arriving in real time. Now, the issue, though, guys, is that Wall Street might just make the same exact mistake they've made with every one of these cycles, or they think that falling cost per task actually results in less AI spend.

Again, there's two sides to this, right? Like, what side do you fall on? You know what side I fall on. It's on the side where AI gets 10x cheaper. I think we're going to use 100x more AI, right?

Dave (2:33)

And we've already seen that. Once Astra examples started getting out there in the marketplace, we saw that people are now just leaving. The tokens are expensive, but they're using less of them, and they're just leaving their computers on and rendering a game for, not rendering, but having Astra build a game for them over five days.

And so it's a different world where, yes, it is more efficient, but it just means we want to leave it running longer and have it doing more for us.

Chris (3:05)

Yeah, Jordan, I want to get your take on this as an engineer. I've been reading every qualified engineer's take I could the last few days on GPT-6 Astra. My one big takeaway is that the big differentiation with GPT-6 is it essentially takes you from A to Z, meaning it will really kind of take over the processes of determining what agents you need, what agents need to be doing at any point in time, and actually making a lot of incremental decisions for you to get you all the way down the path to Z, which is why we're seeing so many really cool zero-shot examples where you just want something and it does it. Now, there's a pro and a con to that, right? The pro to that is it will actually give you a full end result for what you need and make the decisions along that path that it determines to be best, as opposed to waiting and stopping and waiting and stopping, which is a lot of what we see with Anthropix latest AI model. The negative on that is that it's not as great as it relates to doing things ultimately the perfect correct way. It's just going to make a lot of guesses along the way. And so I think a lot of people are having issues with Astra really in terms of controlling it in a defined manner, and working with it, you have to really break it out into little pieces. Otherwise, it does too much. Like it does too much. So I think there's a learning curve here. But as an infrastructure trade, what's most exciting about this, guys, is it just goes nuts with agentic AI. Like yes, the compute is meaningfully down. Like it can do like 3X more with the same compute, roughly is what it feels like. But the amount of compute it's willing to take on to do stuff, it will just run with it. So it feels like one of these scenarios where we're entering a new age of AI that is just going to consume a tremendous amount of compute more than we even thought was imaginable and open up new use cases. But it's not there yet. Guys, we're not there yet. There's so much refinement that is needed on these models that the infrastructure trade just is in the perfect place right now, because we're so intrigued by Astra about what it can do. And yet everything that Astra is showing us it can do is yet opening up more questions to, well, we have an issue here. We have an issue there. It needs to get better here. So it could do so much more. But now we want it to do so much more than that, because we see what it's capable of, but it's not quite capable of giving us a finished product that's actually usable in any meaningful way, okay? We're maintaining a product, keeping up with a product.