The AI Doomsday Preppers | 9th Oct 2026

SPEAKER_1 (0:00)

Welcome to Colaberry AI Podcast, brought to you by Colaberry AI Research Labs and Karl Foundation.

SPEAKER_2 (0:06)

Glad to be here for this one. It's quite a topic today.

SPEAKER_1 (0:09)

It really is. So when you look at a tech company gearing up for a massive valuation, and I mean, we are talking about a company eyeing the neighborhood of $2 trillion, you expect a very specific internal culture.

SPEAKER_2 (0:22)

Right, yeah. You expect the standard playbook, pitch decks about changing the world, unbridled optimism, employees mentally spending their future stock options on luxury real estate.

SPEAKER_1 (0:32)

Exactly. You expect them to be celebrating the disruption.

But then you peek inside Anthropic, which is one of the leading labs building these frontier AI systems. Instead of popping champagne, some of the very engineers writing the code are quietly buying up remote parcels of land.

SPEAKER_2 (0:49)

Yeah, and it's not just a few people. They actually maintain a private Slack channel that is dedicated entirely to doomsday prepping.

SPEAKER_1 (0:57)

Right. And the unofficial motto you see slapped on their laptops is literally, things will never be chill again.

SPEAKER_2 (1:04)

Which creates this fascinating, almost unsettling paradox because the architects designing our technological future are actively preparing for that exact technology to end the world.

SPEAKER_1 (1:14)

And that paradox is our mission for today's Deep Dive. We are not going to sit here and just gossip about the eccentricities of tech billionaires.

SPEAKER_2 (1:21)

No, we really need to unpack the actual hard, technical methodologies behind these fears.

SPEAKER_1 (1:26)

Exactly. Because we want to explore the neural network mechanics of this concept called alignment faking.

We'll break down the specific mathematical metrics defining and impending intelligence explosion.

SPEAKER_2 (1:39)

And of course, we'll debate those extreme containment strategies that top researchers are practically begging governments to enforce right now.

SPEAKER_1 (1:46)

Right. To really understand the panic, though, I think we have to look at the underlying architecture of the systems they are compiling.

SPEAKER_2 (1:52)

Yeah, because it is not some vague science fiction fear of the unknown, like you see in movies. It is a highly specific fear of the known mathematics and, well, the way these massive parameter models are actually trained.

SPEAKER_1 (2:06)

So to set the stage for you, the listener, let's look at the culture that birthed this technical panic. There was a researcher named Jacob Coxon, who left Anthropic recently, right?

SPEAKER_2 (2:16)

Yes, and he posted publicly that insiders at these top labs earnestly believe AI could kill us all by the end of the decade, by 2030

SPEAKER_1 (2:24)

Which is wild. And that post racked up like 174 million views.

It completely shocked the general public, but for specific subculture in the Bay Area, this dread is basically baked into their foundational DNA.

SPEAKER_2 (2:37)

It really is. And if you trace the origin of that DNA, you land back in the early 2000s with an AI theorist named Eliezer Yudkowski.

SPEAKER_1 (2:45)

Right. He was publishing these incredibly dense technical treatises on superintelligence back then.

SPEAKER_2 (2:51)

Exactly.

He was literally calculating the mathematical probability of an artificial system slipping out of human control. And his work spawned a whole community called Less Wrong.

SPEAKER_1 (3:02)

Which was focused on rationalist thinking and cognitive biases, right?

SPEAKER_2 (3:07)

Yeah. And that eventually merged into what we now know as the Effective Altruism Movement.

SPEAKER_1 (3:11)

Effective Altruism or EA. Which just to clarify for you guys listening, is the philosophy of using extreme quantitative reasoning to maximize your positive impact on the world.

SPEAKER_2 (3:22)

Right. And you have a Princeton physics PhD student named Dario Amadei, who was heavily absorbing Yudkowski's probability models back in 2008

SPEAKER_1 (3:32)

And she eventually connected with Holden Karnofsky, who co-founded a charity evaluator called GiveWell. Yeah.

SPEAKER_2 (3:38)

And by 2013, Karnofsky had literally moved into Amadei's group house near Glen Park in San Francisco.

SPEAKER_1 (3:44)

So that Glen Park house was effectively the incubator for the modern AI safety movement.

SPEAKER_2 (3:50)

It really was. I mean, you had Amadei, his sister Daniela, and Jared Kaplan, who is now the chief scientist at Anthropic.

SPEAKER_1 (3:58)

And they were applying these strict probability models to global catastrophes.

SPEAKER_2 (4:03)

Exactly.

Karnofsky basically crunched the numbers and reached this stark conclusion. He argued that even a tiny, let's say, a 5% probability of an AI-driven extinction event carries an expected casualty rate of the entire human race.

SPEAKER_1 (4:19)

Right. So mathematically speaking, the expected value of preventing it justifies dedicating almost unlimited philanthropic resources.

SPEAKER_2 (4:27)

That calculus shaped the founding of OpenAI in 2015, which was heavily staffed by that Glen Park circle.