AI控制问题：代理、成本与机器人

AI is moving from answering questions to taking action, and as it spreads, control is becoming the next big fight. Agents are like teenagers; they're supremely intelligent. They have no fear of fear of any kind of consequence, and you know if you don't put enough guardrails around them, sometimes they might do things that you don't want them to do. There's really a phase we are seeing is scaling into bankruptcy. The United States deployed 1/10 the number of robots that China did last year.

You could get a bill for $50,000 when you thought it was going to be a $20 task. And we, as a country, we've always led these big changes, and we've always been the owner of the next new platform. And I think it's essential if you want to to be a world power that that you own those things, control over what AI does, over what it costs, and in robotics, control over the industries it's about to reshape. Here's that full conversation.

Hi everyone! It is Thursday, July 30th, and the biggest story in AI this week-it's not a new model. It's that once again, money is starting to matter. Since ChatGPT launched, the industry has operated on a pretty generous assumption: spend whatever it takes now, and the returns will come later. But every so often, the market tests that assumption, asks whether revenue is catching up, whether demand is real, whether all that infrastructure will actually earn a return-what kind of return?

That is what we're watching play out again now. A few of the mega caps have reported so far this earnings season, and the message is fairly simple. Google showed growth but was punished for the cost. Microsoft was rewarded for showing clearer AI monetization. Meta was punished because the return on its spending remains harder to see. So the AI trade is volatile these days. Nothing captures that better than what just happened to Leopold Ashenbrenner.

If you don't know Leopold, he is the former AI Open AI researcher who wrote Situational Awareness, basically the definitive Super Bowl case that AI progress is about to go vertical. He then built a hedge fund around that worldview, and it reportedly grew to more than $20 billion in just about two years. Today, we learned that after getting hammered in the recent AI sell-off, the fund sold its public stock portfolio to Citadel.

The Financial Times puts that book at roughly 16 billion dollars and called it one of the largest rushed stock transactions in Wall Street history. Now situational awareness is holding on to its private investments, including an anthropic stake, but the public trade is now in Ken Griffin and Citadel's hands. And in a nice bit of timing, chip stocks they are ripping today. That could change tomorrow, but more importantly, it shows why the Citadel sale doesn't settle anything about the underlying AI thesis.

May tell us more about leverage, timing, liquidity. You can be right about where the technology is going and still get forced out of the trade before it gets there. At the same time, Dwarkes Patel has been making an interesting point about the chips themselves. A lot of the bearish AI math assumes that these GPUs they age like iPhones. Nvidia releases a new generation; the old one becomes less useful, and big tech has to start replacing hundreds of billions of dollars in hardware.

That may be true still, but Dorkhash argues that the software running on those chips keeps improving, so the same NVIDIA GPU can run a better model today than it could say a year ago. Inference gets more efficient, the output gets more useful, the cost of completing a task comes down. So the hardware may get older while its earning power goes up, and that is the bull case, essentially a super cycle. But the money around this is getting a little more complicated.

NVIDIA may help finance OpenAI's data centers and the NVIDIA chips inside them. And today, OpenAI cut the price of two new models-one by 20 percent, the other by 80 percent-just three weeks after launch. That could mean the cost curve is working. Could also mean customers are pushing back, and pricing power is getting weaker. So the question isn't whether AI is real; it's whether the returns can arrive before investors and customers lose their patience.

Is AI's easy money era ending? To get past the stock charts and into what is actually happening underneath the AI trade, we are joined by three people building at different parts of the stack. We've got Cisco's G2 Patel right beside me on the Open AI Rogue Agent AI Security and whether autonomous systems need a kill switch. We've also got Fireworks AI CL Lin Chiao on OpenAI's price cuts, the model price warn, whether open models have an advantage as customers focused on cost, and then finally we're going to talk to Standard Bot CEO Evan Beard on new restrictions targeting Chinese robot imports and what they could mean for the cost of automating American factories, let's start with g2.

Welcome. I love how you're willing to come into the studio. It's great to have this. You know, I've I've actually been in San Francisco three days in a row. It's kind of I think I should like move here or something because everything's happening. All the alphas happening in SF. I mean, I've been on vacation for the last two weeks, and I you know can't. I know. How was it? You have a fun time. This is great, but I'm happy to be back.

X missed you. You know, two weeks in AI is like two years. I'm stealing that from someone I can't remember who, but that's what it feels like. Do you feel completely out of touch? Totally, and so you're here to help me get back in touch. But open source happened. Open source happened huge. Broke out into the open, although you know we've been talking about it forever, and it feels like this sort of-it's always here-the bull bear debate.

But you know, where are we? That's like where I'd like to start in the AI cycle right now. I mentioned Dorkach. I could see you kind of nodding along. Did I represent sort of his argument? Yeah, there was one piece that was actually-we'll talk about Dwarkash in a second. But if you take a step back and say what is actually happening, there's not that much that, in my mind, changes in the macro picture in the sense that there is a secular shift.

Secular shifts will have, you know, constant levels of ups and downs, you know, from week to week. But the fundamental assumptions of the fact that AI actually will help help us solve problems that we couldn't solve, we couldn't think of solving before AI. I think still remains, and so I tend to not get super kind of fixated on what the weekly movements are, and I actually focus more on what are the big kind of shifts that are occurring in in intelligence, and how can we make sure that we continue to keep taking advantage of them, and and so that that to me is exciting.

Now, what what Dwarkes talked about is, by the way, that that boy has a lot of alpha. He's like you know he's very very bright. And one of the things that he he brought up, which was very cool, was this. And this is an important thing. I've actually talked about quite a bit, which is your specifically with chips. This is the first time that we've seen where these chips continue to keep, you know, having longer and longer shelf lives as you actually see the compression of time for a new chip getting developed, which is very counterintuitive because usually if it takes your chip development cycle goes down, you would think so would your useful use of that chip and the lifetime value of the chip go down.

But no, that actually tends to go up, and because we have we are in a constrained environment, that actually will continue in the you know for the foreseeable future. And that's like that's essentially the super cycle that people have talked about, and Dworkash put it really well. Yeah. But like, let me push back on that a little bit. I remember Jasmine and I did a story a while ago about you know the data center that Oracle was building, an

原文超出正文长度上限，此处截断——上游还有内容，完整版见上方「原文 ↗」。