Jensen's in Seoul pitching a South Korea AI bet — while the rest of the podcast world argues over who's allowed to copy his customers' models. Quick catch-up before we dig in: Kimi K3's release put Chinese open-weight AI back in the strategic spotlight. Then Xi Jinping's World AI Conference speech made it a governance question, not just a model-release story. So does Beijing keep backing open source, rein in frontier releases, or move toward tighter state control? This is the Tech Podcast Podcast — today, Jensen and Satya taking very different CEO laps, All-In fighting over open source, and Linear Digressions quietly explaining the mechanism they're all talking about. Let's start in Korea. If Chinese open-weight AI governance matters to you, hit follow — we'll be back on it soon. Here's Bloomberg Tech:
In a special bonus episode of Bloomberg Tech, host Ed Ludlow interviews Nvidia CEO Jensen Huang following his appearance at a Korean AI Summit and meeting with South Korean President Lee Jae Myung. Huang announces major partnerships in South Korea, including a $500 billion business relationship with SK Group covering memory purchases and AI supercomputer sales, a $1 billion investment in Naver, and plans to build over 2 gigawatts of AI data centers.
Fourteen minutes and forty-two seconds. Jensen walks off a Korean AI Summit stage, and Ed Ludlow gets a fourteen-minute window to make it count. And the headline number is a five hundred billion dollar business relationship with SK Group. A billion into Naver almost reads like a rounding error next to it. I want to unpack the two gigawatts of data centers in South Korea. Why there? That's what I'd hammer on — what does Seoul offer that the US-China corridor doesn't right now? Right. Did Ludlow ask that, or did he let Jensen run the goodwill lap? Because the tenfold-semiconductor-growth line is a great applause moment, but it tells you nothing about why he picked this partner. What caught me was Jensen making the case for open models — safety and innovation — on a stage where he's selling the compute either way. It's hardly a neutral position. The guy selling the shovels wants everyone digging. Hold that thought, because open models come up again in a minute. This one's from Digg:
The All-In Podcast returned with hosts discussing the battle to preserve open source AI against regulatory pressures from companies like Anthropic and OpenAI. Topics include Chinese AI distillation strategies and an Anthropic settlement. David Sacks and Chamath Palihapitiya highlighted the episode, noting its focus on open source preservation amid industry shifts.
So the besties are back, and they're framing it as "the fight to save open source AI" — against Anthropic and OpenAI doing regulatory capture. Big words. I wanna know if Sacks and Chamath actually name a mechanism or just wave at the Kimi K3 panic. It's the U.S. ban debate we flagged when Kimi K3 first spooked everyone, only now it's center stage on a mainstream pod. And they've got the Anthropic settlement to hang it on — one-point-five billion, one of the largest resolutions tied to this whole AI IP mess. Chamath calls it "the great IP theft hypocrisy." Great line, but I want the argument underneath it. Right, because the rundown leans hard on "Chinese distillation strategies" and never says what distillation actually is. Convenient, since we've got a whole segment coming that spells out the mechanism. Exactly. You can rail against distillation as theft for forty minutes and still dodge whether hammering an API for training data is what you're actually mad about. Fareed Zakaria GPS, with Fareed Zakaria:
Fareed Zakaria hosts an exclusive interview with Microsoft CEO Satya Nadella, discussing AI's economic impact, competition with Chinese models, data control, and the future of work. The episode also features an interview with Venezuelan opposition leader María Corina Machado, who remains in exile after the capture of Nicolás Maduro, and a closing segment analysing Iran's economic resilience during war.
So we just heard Jensen do the investment-goodwill lap in Korea, all upside — and then Fareed gets Satya for 42 minutes, and the tag on the episode is literally 'sombre.' Same week, two Microsoft-adjacent AI titans, totally opposite register. Yeah, and Nadella earns that tone. His line is that AI has to drive broad economic growth or it's a bubble. He's putting a failure condition on his own hype. What made me sit up was the 'reverse information paradox' — AI providers learning from the customers who use them. He's naming the exact thing every enterprise founder is scared of. Right — a 26-year-old building on this stuff wants to know if the big lab walks in and eats their market. Nadella basically confirms how that happens, then pivots to wanting legal and technical protections for firms. He sees the threat; he's not pretending otherwise. And on jobs, he doesn't do the usual 'net positive, don't worry' dodge — he says new jobs, new value, but real displacement in the transition. For a CEO, admitting the transition hurts is practically confessional. He's the one CEO this week actually saying where the value goes instead of just gesturing at it. Whether the protections he's floating are real or wishful is why I'd queue the full 42. Linear Digressions writes:
They cover the two big reasons labs do this — making lighter, faster, more focused models for specific tasks, and the more contentious use case of effectively copying a rival's flagship model by hammering its API with questions (with a callback to the old Bing/Google search controversy).
Okay, this is the episode that finally names the mechanism All-In spent the whole show dancing around. Distillation — you point a small student model at a big teacher's API, hammer it with questions, and train on the answers. It's what everyone means by copying, but they keep gesturing at it without saying the word. And there's a clean split. One use is legitimate — a lighter, faster model for a narrow task. The other is basically photocopying a rival's flagship through its own front door. Best part: some of these student models introduce themselves as "Claude." You literally catch them wearing the other model's name tag. What gets me is how old the idea is. Hinton wrote the soft-label paper in 2015 with Jeff Dean and Oriol Vinyals — the full probability distribution, not just the top answer. A decade-old research trick is now something labs want a regulator to ban. Proving it is nearly impossible. There's no watermark on a probability distribution. It's the Bing-scraping-Google fight again, just with weights instead of search rankings. That answers the scale question too — a smaller model closes the gap because it's trained on the big one's outputs, not from scratch. The catch is, it inherits the confident wrong answers too. Right — you copy the genius and the hallucinations in one pass. So when Jensen's out there pitching physical-world AI and Satya's doing 42 sombre minutes on the economy, this is the plumbing nobody on those press tours wants to explain. Have feedback, a story idea, or a correction? Email us at techpodcastpodcast at lantern podcasts dot com. We’d love to hear from you.
You’ll find links to every story in today’s show notes, so if something caught your interest, you can read more there.
That’s Tech Podcast Podcast for today. This is a Lantern Podcast.