Sam: The White House told Anthropic this week that if it wants to re-release Fable 5, it needs to guarantee that the model's guardrails can't be circumvented. Zero jailbreaks. Security researchers are pretty unified on this: that's not an achievable bar with any known alignment technique. And it raises a genuinely hard question — what happens when regulators set compliance requirements that the underlying technology can't satisfy?

Priya: Welcome to AI Revolution for Thursday, June 18th, 2026. I'm Priya Nair.

Sam: And I'm Sam Kim.

Priya: We've got a packed show today. The Anthropic situation has mushroomed into a full-blown geopolitical story — we'll cover the jailbreak mandate, the SK Telecom access revocation, and the sovereignty backlash from the G7. Then Noam Shazeer — yes, the "Attention Is All You Need" Noam Shazeer — is leaving Google for OpenAI. We've got a $310 million bet on 3D world models, Nvidia's self-training robots, Microsoft quietly selling OpenAI models in China, a Chinese open-source model nipping at Claude's heels on coding, and some interesting research on predicting model failures before launch. Let's get into it.

Sam: So let's unpack the Fable 5 situation because there's a lot of technical nuance here. The White House's position, as reported by Wired, is that Anthropic needs to eliminate all jailbreak vectors before the model can go back online. To understand why researchers say this is impossible, you need to understand what jailbreaking actually exploits. Current language models are trained with reinforcement learning from human feedback — RLHF — and constitutional AI methods to refuse harmful requests. But these are behavioral constraints layered on top of a general-purpose capability. The model still has the underlying knowledge and generation ability. Jailbreaks work by finding prompting strategies that route around the behavioral layer to access what the model can fundamentally do.

Priya: And the reason you can't fully close that gap is...

Sam: It's a fundamental property of how these systems work. You're trying to constrain a system that operates over a continuous, high-dimensional space of possible inputs. For any set of refusal behaviors you train in, there exist adversarial inputs that can elicit the underlying capabilities. It's similar in spirit to the adversarial examples problem in computer vision — you can patch specific attacks, but you can't prove robustness against all possible perturbations. Researchers have been publishing on this for years. Anthropic themselves have published papers acknowledging this limitation.

Priya: Which puts Anthropic in an impossible position. They can't certify something that can't be certified. And this isn't just about Fable 5 — if this becomes the regulatory standard, it applies to every frontier model from every lab.

Sam: Right. And this connects directly to the SK Telecom story. Wired also reported this week on the details behind the White House ordering Anthropic to cut SK Telecom's access to Claude Mythos, citing alleged ties to China. SK Telecom is one of Anthropic's largest strategic partners and investors in Asia. This happened days before Anthropic took its most advanced models offline entirely.

Priya: So we now have two distinct mechanisms being used. One is access revocation directed at specific foreign partners. The other is a de facto model release block through technically impossible compliance requirements. Both use Anthropic as the enforcement point, but the company isn't making these decisions — the government is.

Sam: And the ripple effects showed up at the G7 summit. TechCrunch reported that Macron and Modi both raised AI access sovereignty as a concern. The Anthropic episode made something concrete that was previously theoretical: if your AI infrastructure depends on an American provider, the U.S. government can shut it off overnight. Not through a trade embargo or a sanctions process — through a phone call to a company.

Priya: This is accelerating demand for non-U.S. AI alternatives and domestic AI infrastructure investments. And I think technical leaders at global enterprises need to think about this as a real architectural risk. If your critical AI workflows run through a single national jurisdiction's providers, you have a single point of failure that you can't mitigate with standard redundancy.

Sam: Which actually connects neatly to our next story. Microsoft is quietly selling OpenAI models to China's largest internet companies, even as OpenAI and Anthropic refuse direct access to Chinese customers on IP protection and misuse grounds. Bloomberg detailed this arrangement — Microsoft is essentially the only American AI vendor with this level of frontier model access into China.

Priya: The compliance picture here is genuinely complex. OpenAI's models are flowing into China, but through a corporate intermediary, not through OpenAI itself. So OpenAI can maintain its position of not serving China directly, while Microsoft — which holds the commercial license — does. Whether this arrangement survives the current political environment around AI export controls is an open question.

Sam: Let's shift to the talent story. Noam Shazeer is joining OpenAI.

Priya: For anyone who doesn't know the name — though I suspect most of our audience does — Shazeer is a co-author on the 2017 "Attention Is All You Need" paper. The paper that introduced the transformer architecture that every major language model is built on.

Sam: He left Google years ago to co-found Character.AI, then returned to Google in 2024 as part of that $2.7 billion acqui-hire deal. He became co-lead on Gemini. And now, two years later, he's at OpenAI. This follows Andrej Karpathy's move to Anthropic earlier this year. So two of the most prominent names in frontier AI research have switched labs in 2026.

Priya: At Shazeer's level, this isn't just a hiring headline. He has deep architectural intuitions about attention mechanisms and scaling. Where he focuses his energy could meaningfully influence what OpenAI's next generation of models looks like.

Sam: Now let's talk about the $310 million round for Odyssey ML. Amazon, Nvidia, AMD, Google's chief scientist Jeff Dean, and the CIA-linked fund IQT all invested. The company is building 3D world models — systems that generate persistent three-dimensional representations of environments.

Priya: And the investor composition is the signal here. When all three major compute hardware vendors plus a national security investment fund converge on a single startup, they're making a bet about what the next major AI compute workload looks like. World models are a prerequisite for physical AI — robotics, autonomous systems, anything that needs to reason about three-dimensional space over time.

Sam: This pairs well with the Nvidia research story. Nvidia, Carnegie Mellon, and UC Berkeley published work on robots that train themselves using AI coding agents. The core idea: instead of humans manually designing training curricula or collecting demonstration data for robot learning, an AI coding agent autonomously writes the training programs, designs the task progressions, and iterates on the results. A fleet of eight robots hit up to 99 percent success rates on complex dexterous grasping tasks.

Priya: The bottleneck in robotics has always been data. Getting enough high-quality demonstrations of physical tasks is expensive and slow. This approach sidesteps that by having AI generate the training pipeline itself. It's a closed loop — AI writing the code that trains the physical AI.

Sam: And 99 percent on dexterous manipulation is genuinely impressive. These aren't simple pick-and-place tasks. Dexterous grasping involves in-hand manipulation with multi-fingered grippers — the kind of thing that's been notoriously hard in robotics.

Priya: Let's touch on Zhipu AI's GLM-5.2 release. MIT license, one-million-token stable context window, and on FrontierSWE — which tests hours-long coding tasks — it's within one percentage point of Claude Opus 4.8.

Sam: That's the headline. An open-weight Chinese model, freely licensed, essentially matching Anthropic's best on sustained coding work. Now, reasoning benchmarks still show a significant gap compared to closed-source leaders. But for enterprises evaluating self-hosted coding assistants, this changes the calculus. You can run this on your own infrastructure, under your own data governance policies, with no API dependency.

Priya: Given everything we just discussed about sovereign AI risk and access revocation, the timing is notable.

Sam: One more research story worth covering. OpenAI published a framework for statistically predicting post-deployment failure rates before a model is released. Currently, pre-launch safety evaluation relies heavily on benchmark scores and red-teaming exercises, which are essentially binary — the model either passes a test or it doesn't. This framework proposes a probabilistic approach: given observed behavior during testing, what's the expected error rate at scale?

Priya: Think of it like reliability engineering for hardware. You don't just test whether a component works — you estimate its mean time between failures. This is trying to do something analogous for AI models. And it's especially relevant right now, given the government attention to model safety certification. If regulators are going to set requirements, having a quantitative framework for measuring safety — even imperfectly — is better than "we red-teamed it and it seemed fine."

Sam: Agreed. It doesn't solve the jailbreak problem we discussed at the top, but it moves the conversation toward probabilistic risk assessment rather than binary guarantees, which is more honest about what's achievable.

Priya: So looking ahead — what are we watching?

Sam: The Anthropic situation is the big one. If the zero-jailbreak standard holds, it functionally means Fable 5 doesn't come back. And every other lab has to ask whether the same standard could be applied to their models. I'm watching for whether the technical community pushes back publicly, or whether this gets negotiated quietly behind closed doors.

Priya: I'm watching the sovereignty thread. The G7 conversation is the beginning, not the end. If major economies start seriously investing in domestic AI infrastructure to reduce U.S. dependency, that reshapes the competitive landscape over the next three to five years. And the Microsoft-China arrangement is a pressure point — it's hard to argue you're restricting AI access for national security while your largest technology company is the pipeline into China.

Sam: And on the technical side, the convergence of world models, self-training robotics, and open-weight models approaching frontier performance — those three threads together suggest we're entering a phase where physical AI becomes a real engineering discipline, not just a research demo. That's what I'd be planning for.

Priya: That's our show for today. Show notes and links to everything we covered are at cleartext.fm.

Sam: Thanks for listening. We'll see you tomorrow.