TCG084: Is AI Going to Kill Us All? Separating Frontier Risk From Frontier Theater

William (0:12)

Is AI going to kill all of us? Is the Terminator coming for you?

Some of the people building it say there's a real chance it will. They're asking companies to slow down development. They're asking governments to intervene, and they're asking the public to trust their judgment. Their judgment, because, hey, it's never been wrong before, or has it? Before we hand anyone that kind of power, we believe that it's worth examining the evidence, the incentives and the track record of some of the big players that are peddling some of this stuff that you see in the media. Evonne is back with me for this one, just us two. There has been so much spicy news that it's like, we just can't help ourselves. So how are you doing today, Evonne?

Eyvonne (1:14)

I am doing great. It has been a very busy week of business travel. We are both in the thick of sports and kids activities. If fall is here, even if the calendar doesn't indicate it yet.

But interestingly enough, the heat in our part of the world has still not relented.

William (1:35)

It is hot. I took a walk yesterday. I'm coming off a broken ankle, so it's a little slower for me. So I took a walk. I made it about a half a mile in the blazing heat, by the way.

It was hot. Very humid, and I regretted all my life's choices for deciding to take a walk yesterday because it was hot.

Eyvonne (2:00)

Yeah, we had a band competition and two back-to-back soccer games yesterday. And the thermostat said 84 degrees, but it felt a lot warmer sitting out there in the blazing sun. So yeah.

William (2:14)

I believe that, yeah, with me it's cross-country, so outside very hot.

Eyvonne (2:19)

Which is the most boring thing for parents is cross-country. We've done cross-country and blessed your heart, that is a labor of love to watch a cross-country race because you're just standing there.

William (2:31)

Yeah, and the traffic that they jam into like some park or some place that wasn't meant to have that much traffic, it's undeniably, it's awful. It's crazy, but yeah. Onto that and then to ice hockey for me. But yeah, let's get to the episode. I think this is going to be an interesting one. There's a lot of stuff that's happened and I just want to preface this by saying, look, I might get a little spicy here, a little of my opinions might come out with this. You might feel it, maybe you won't. But today we're really going to kick this off with, honestly trying to frame the story and check, I'd say, historical predictions.

And after that, we'll look at the, I think, the campaign around the recent Jacob Coxon's resignation from Anthropic. And then Dario's proposal to slow down frontier development and the financial things, the contradiction that's maybe facing Anthropic at this point. And I'm going to try and frame each part. Like, as you know, Evonne, I've been texting you about this. I've been fired up about this. I normally don't get fired up about this stuff, but this whole thing has got me very annoyed. You're triggered. And I don't get triggered like this. Look, there's a thing called work and life balance. On the work side, it's like, hey, I do what I can and then leave the rest. But this thing has me like bonkers right now, because it's just so bizarre and so crazy. And shame on the media for just propagating and promoting stuff that they haven't done. They're journalistic, they're journalists. They should be doing their research. They should be fact checking. They should be doing some sort of investigation instead of just promoting stuff. But that's why you have the small contingent of technical podcasts with technical folks that have actually worked inside of these with the technology to actually come and try to frame the argument or frame a different perspective. So the person, I would say, that's at the center of all of this, like I said, Jason Erden, I keep wanting to say Jason, Jacob Coxon, like under 30-year-old British AI researcher with a heavy mathematics background, I think from Cambridge. So he worked on, I think, primarily pre-training at OpenAI and Anthropic.

And I think he was a core contributor in the early phases of, I think GPT-4, GPT-45, I don't know, I can't remember, but he co-authored some research on like mechanistic interpretability things, because everybody knows what that is. So on the other day, after about four months of working at Anthropic, and hey, four months is so long, a lot of funny, I've been listening to some funny podcasts that were like, hey, four months is about the time it takes to find the restroom and get your laptop set up with all those security agents.