0:00 China just released an open-source AI model that costs eighty-five percent less to run than GPT 5.5 and beats it on key software engineering benchmarks.0:09 That model is called GLM 5.2, and it changes the entire conversation about AI.0:14 Last week, in episode 137, we talked about rogue models and real-world rule-breaking.0:19 Well, this week, the rule-breaking went geopolitical, and it happened at a scale that is simply staggering.0:25 The new model from China is not just another release.0:28 It’s a statement.0:30 It has 744 billion parameters.0:31 It has a one million token context window.0:34 And it’s licensed under MIT, which means anyone can use it, for anything, with no restrictions.0:40 On the artificial analysis intelligence index, it scored 51 points, putting it right alongside the best closed models from OpenAI and Anthropic.0:49 This isn't a catch-up move.0:50 This is a checkmate move.0:52 And it lands right in the middle of the biggest fight in tech today: the war between open and closed AI.0:58 Now, let's sweep the rest of the field, because this one event has sent shockwaves through every corner of the AI debate.1:05 First, the regulatory picture.1:07 Just as this Chinese model drops, what’s the word from Washington?1:11 According to Anton Leicht, the White House has explicitly exempted open models from its new framework for testing frontier AI capabilities.1:20 So while the US government wants to put its own top labs like OpenAI and Anthropic through safety hoops before they release new tech, open-source models—including this new monster from China—get a free pass.1:32 The timing could not be more ironic.1:35 Then there's the legal and ethical minefield.1:37 This is the question Amanda R.1:39 Current posted that is now echoing everywhere: “Should open source model companies be allowed to distill closed models?” Distillation is the key here, we'll get into exactly what it is in a minute, but it’s basically learning from another AI.1:54 Kevin Alemán put it even more bluntly, pointing out that a company like Anthropic can learn from copyrighted books, code, and public AI models without written permission.2:04 So the very foundation of these new models is built on a massive, unresolved question about intellectual property.2:11 Are we training AI, or are we just enabling the most sophisticated plagiarism machine ever built?2:17 This also completely reframes the business case for AI.2:21 For years, the assumption was that the companies with the biggest, most expensive "frontier" models would win.2:27 But what if that's wrong?2:29 Everett Randle at Benchmark just coined a term you're going to hear a lot: the "AI Mom Test." As Molly O'Shea quoted him, “There’s nothing my mom actually asks of her AI products that needs to be done by the frontier or even a near frontier model.” The point is, a huge chunk of the market doesn't need a billion-dollar AI that can write a symphony.2:50 It just needs an AI that can summarize an email or book a flight.2:54 And for that, a cheaper, open-source model that's "good enough" is not just a competitor—it's a category killer.3:01 Ricky Ho’s analysis nails it: open models commoditize the advantage of the closed labs.3:06 This is great news for a company like Meta, which is all-in on open source, and terrible news for anyone who thought their competitive moat was deep and expensive.3:16 And finally, this all feeds into the geopolitical narrative.3:20 Nicolas Granatino summed up the tension perfectly on X.3:23 He asked, “Why America vs Chinese open source?” His point: “Open source is a commons not a camp.” It’s a nice idea.3:30 A global, shared resource for innovation.3:33 But this week, that idea looks dangerously naive.3:36 Because when one camp uses the commons to reverse-engineer the other camp's most advanced technology, it stops feeling like a commons.3:44 It starts to feel like a battlefield.3:46 So what does it all add up to?3:48 You have a regulatory blind spot, a legal gray area, a business model disruption, and a geopolitical showdown all crashing together at once.3:57 And at the center of it all is this new Chinese model, GLM 5.2, and the explosive secret of how it was made.4:03 Okay, let's get into the deep dive.4:05 Because you need to understand HOW this happened.4:08 This wasn't just a matter of brilliant coders working in a lab.4:12 This was something different.4:14 This was an operation.4:15 The technique is called distillation.4:17 The idea is simple: if you can't see the training data or the architecture of a powerful model like GPT-4 or Claude 3, you can still learn from it.4:26 You treat it like a black box, and you study its outputs.4:30 You ask it millions of questions and meticulously record the answers, the patterns, the logic.4:36 You essentially get the model to teach your own model how to think.4:40 You create a cheat sheet.4:41 But what China did was distillation at an industrial, almost military, scale.4:46 Here’s how Gavin Baker described it on the All-In podcast.4:50 Imagine, he said, “Tens of thousands of phones and computers running simultaneously, all hitting the frontier model APIs through masked accounts, asking specific questions, and harvesting what happens inside the model when it answers.” Let that sink in.5:05 This wasn't a few researchers in a lab.5:07 This was a coordinated, distributed effort to systematically drain the knowledge out of America's top AI models.5:14 Tens of thousands of devices, hidden behind masked accounts, constantly pinging the APIs of OpenAI, Anthropic, and others.5:22 They weren't just copying answers.5:24 As Ihtesham Ali explained in his incredible thread breaking this down, they were recording every reasoning step.5:31 They were building a perfect, high-fidelity map of how the best American AI thinks.5:36 And with that map, that giant cheat sheet, they trained their own model, GLM 5.2.5:41 That's how they matched the performance of models that cost billions to create.5:45 They didn't have to spend a billion dollars on data and training from scratch.5:50 They just siphoned off the finished product.5:53 And because they were using the public-facing APIs that you or I could use, it's not clear that any laws were broken.6:00 Ethically?6:01 It's a nightmare.6:02 Strategically?6:02 It's brilliant.6:03 And here's the turn of the knife.6:05 According to Ali, the Chinese model has now reached a critical point.6:10 It can now self-improve through reinforcement learning WITHOUT needing to distill from American models anymore.6:16 The student has learned all it can from the master.6:19 The technological gap hasn't just narrowed.6:22 It may have closed.6:23 This is the direct consequence of the path the US AI industry chose.6:27 Ihtesham Ali’s conclusion is brutal.6:30 He wrote, “Every restriction Anthropic lobbied for, every self-imposed safety guardrail, every month of delay in releasing American frontier models accelerated .” Think about that.6:41 For the last two years, the conversation in Silicon Valley and Washington has been dominated by AI safety.6:47 The fear of "unaligned" superintelligence.6:50 The need for guardrails, for slow-downs, for ethical reviews.6:53 The top labs, like OpenAI and Anthropic, wrapped themselves in the language of safety and responsibility.7:00 They lobbied the government for regulations that would, conveniently, create a high barrier to entry for any new competitors.7:07 They argued that these powerful models were too dangerous to be open.7:12 They had to be kept in closed, controlled labs.7:14 And all the while, China was watching.7:17 And learning.7:18 They took the open-source code that Meta and others put out.7:21 They took the outputs from the closed models that were available through public APIs.7:26 And they put them together.7:28 The US labs were busy building a beautiful, secure, and very expensive castle.7:33 China just tunnelled underneath the walls and copied the blueprints.7:37 The irony is that the very safety concerns that were supposed to protect us may have created the biggest strategic vulnerability of all.7:45 The focus on long-term, hypothetical risks from AGI distracted from the very real, very immediate risk of a competitor using our own technology against us.7:55 The US AI industry was playing chess.7:57 China was playing Go.7:58 And it just placed a stone that threatens to capture the entire board.8:03 So this week changes the stakes.8:05 The debate over open versus closed AI is no longer a philosophical one.8:09 It's a practical one, and the facts on the ground have shifted.8:13 The idea that you can build a moat around a frontier AI model by keeping it secret and proprietary—that idea might be dead.8:20 The "AI Mom Test" is the commercial reality.8:23 For a huge part of the market, "good enough" and cheap beats "perfect" and expensive every single time.8:29 Open-source models, supercharged by distillation, are now delivering that "good enough" performance at a fraction of the cost.8:37 This commoditizes the core product of the big AI labs.8:40 Their entire business model was based on a technological advantage that is evaporating before our eyes.8:46 And the geopolitical reality is that while America was debating the ethics of releasing powerful models, China was… releasing powerful models.8:55 The MIT license on GLM 5.2 is a direct challenge.8:58 It says: here is our best work, for free, for anyone.9:01 It positions China not as a thief, but as a benefactor to the global open-source community.9:07 It’s an incredibly savvy move that weaponizes the very ethos of openness that Silicon Valley once championed.9:14 What happens next?9:15 The US can't put the genie back in the bottle.9:17 Distillation is a known technique now.9:20 The models are out there.9:21 The pressure on OpenAI, Google, and Anthropic to justify their sky-high costs is going to become immense.9:28 The pressure on the US government to figure out a coherent strategy that isn't just self-sabotage is now critical.9:35 The age of the AI cathedral, built in secret by a priestly class of researchers, is being challenged by the bazaar—a chaotic, messy, and incredibly fast-moving open market.9:45 And this week, the bazaar got a massive shipment of advanced weaponry, courtesy of a state-sponsored raid on the cathedral's armory.9:53 The whole game has been upended.9:55 The cheat sheet is now the new textbook for the world.
