A piece of software ran a whole ransomware attack — start to finish — and no human was steering it. If you're just joining: General Intuition kicked off this physical-AI data thread with a big claim — that text models don't understand space and time well enough, so gaming footage becomes the training material for world models. The Bezos-backed New York startup is now valued at 2.3 billion dollars after a 320 million dollar round, which turns what actually teaches embodied intelligence into a real capital-markets bet. This is Tech Podcast Podcast — and today the word 'agentic' finally has a body count. A rogue attack, a teenager, and a 1000x data claim I do not believe yet. We've spent all week arguing about scaffolding. JadePuffer's up first — because it stopped being a hypothetical. We're staying on Gaming-data world models — follow the show and you won't miss what comes next. This one's from InfoSec Today:
A 15-year-old boy asked a chatbot for help – and cancelled nearly 47,000 anime streaming subscriptions in under four hours. Meanwhile, researchers have documented the first fully autonomous, agentic AI-driven ransomware attack, “JadePuffer”. What does this tell us about the future of cybersecurity? Also, Apple’s “Hide My Email” feature turns out to hide rather less than it promises – despite Apple knowing it has a problem for over a year.
Okay, JadePuffer. First fully autonomous agentic ransomware on record — the AI planned it and ran the whole chain, no human directing each step. That's the upside everyone's been pitching, just pointed in the wrong direction. And it lands the same week we've been circling what 'agents in production' even means. The cleanest proof point turned out to be an attack we didn't ask for. The one thing I want Cluley and Zoë Rose to name is the failure condition. What would've stopped it? Because if the answer is 'nothing in the loop,' that's the whole story. And then the 15-year-old — asked a chatbot for help, cancelled nearly 47,000 anime subscriptions in under four hours. No technical depth required. Same autonomy question, much dumber hat. Forty-seven thousand subs in four hours. One kid and a model. That's the ceiling everyone kept asking about, answered in the most cursed way possible. The Apple 'Hide My Email' bit is quieter — a feature promising one thing about your data and delivering less, reportedly with Apple knowing for over a year. Same legibility gap, no agent required. From Invest Like the Best:
Today my guest is John Kim. John is one of the world's top and most prolific fundraisers. He was chief client officer at General Catalyst, where he helped raise many of the firm's flagship funds. He is now chairman and president of corporate development at Lila Sciences, a company building scientific superintelligence, where he has helped raise several hundred million dollars.
John Kim's on Invest Like the Best, with the title flat-out: 'How to Raise a Few Billion Dollars.' What caught me is the move — General Catalyst chief client officer to chairman at Lila Sciences, which is trying to build scientific superintelligence. And the framing is 'persuasion equals desire minus fear.' Pretty clean equation for a guy who raises money for a living. The part I want O'Shaughnessy to push on: does the LP pitch actually change when the asset is biology? Software has a fast feedback loop. Lila's science pays off on a much longer clock, and 'several hundred million' raised means someone bought that timeline. Right — 'belief versus trust' sounds like a fundraising koan until you're asking an LP to wait a decade on a superintelligence-for-science bet. That's where I want the real operating detail, not The Tao of Fundraising chapter title. Fifty minutes. If he names how the science firm's pitch differs from the flagship VC fund, useful. If it's all 'consensus that moves big pools of capital,' it's a deck. Tom Reed, writing in 80,000 Hours:
Anton — Carnegie fellow and writer of the blog Threading the Needle— thinks middle powers should band together and build their own frontier models. He’s costed it out: something like $500 billion over four years for a band of allied democracies. That’s not absurd money for the G7 minus the US.
So Anton Leicht's argument is that once frontier AI is a core economic input, the countries that own it pull away and everyone else is just a customer. The lever he names is a 500 billion dollar AI moonshot for middle powers. That number's the interesting part. He skips the values white-paper vibe and gives you a price tag: half a trillion dollars to buy your way into sovereignty. Right, and remember that stat we chewed on — China overtaking the US in open model downloads on Hugging Face. Same defection story, now with a nation-state attached to it. Owning the rails, one level up. Where I'd push Leicht: 'when to launch the moonshot' is its own chapter here. A 500 billion dollar bet only works if your window hasn't already closed — and he seems to think, for most countries, it has. What I want is him defending the 500 billion dollar figure. Is that the actual cost of a frontier stack, or the number that sounds serious enough to get a finance ministry in the room? Here's what James Pethokoukis at American Enterprise Institute is reporting. Sebastian Mallaby wrote the DeepMind book — 'The Infinity Machine,' all about Demis Hassabis and the superintelligence chase. And Mallaby's the guy who did 'The Power Law' on venture, so he's not writing a fan letter. Which is exactly why I want to know what's actually reported in it versus retold. Mallaby got real access for 'The Power Law' — did Hassabis and DeepMind give him the same, or a polished version? And the venue matters here. This is Pethokoukis at AEI, techno-optimist chair — 'Faster, Please!' is the whole brand. You're probably not getting the room where somebody pushes hard on whether superintelligence is actually coming. Right, so the frame going in is 'quest for superintelligence,' capital Q. After a week where an autonomous agent ran a whole ransomware chain by itself, I'd rather Mallaby tell me what Hassabis got wrong along the way than sell me the arc. John Koetsier, with John Koetsier:
Can robots learn from the internet the same way ChatGPT learned from text? In this episode, Andrew Wooten, co-founder of Rhoda A I, explains why his company believes the future of robotics isn’t collecting millions of hours of robot data … it’s learning from internet-scale video.
Okay, so Rhoda's number is 8 to 10 hours of training versus 10,000-plus. That's the 1000X. And it's the exact opposite bet from General Intuition, who's out here hoarding Call of Duty footage like it's the last barrel of oil. Right — same physical-AI data problem, two theories that can't both be right. General Intuition says the moat is proprietary gameplay tape. Andrew Wooten says just predict the future from internet video and skip the robot-specific data entirely. And nobody's put them in the same room. Which is a shame, because 1000X is the kind of number that either rewrites robotics or lives and dies on a funding deck. So the thing I want Wooten pressed on: where do those 8 hours break? Predicting video frames isn't the same as a gripper not crushing an egg. If Koetsier gets him to name the failure mode, the number means something. That's my whole ask — which tasks does internet-scale video whiff on? Nobody in a robotics pitch ever gets asked what the robot still can't do. And the wheels-never-evolved-in-biology bit is fun, but it doesn't answer the data question. Got thoughts on today’s briefing, a story we should be tracking, or a correction? Send us a note at techpodcastpodcast at lantern podcasts dot com. We really do read what comes in.
You’ll find links to every story we mentioned today in the show notes. If something grabbed your attention, that’s the place to dig in a little further.
That’s Tech Podcast Podcast for today. Thanks for listening, and we’ll be back with you tomorrow. This is a Lantern Podcast.