Sam: Anthropic put out a statement yesterday that's worth taking seriously. They say they're approaching what's called recursive self-improvement — RSI — the threshold where an AI system can meaningfully contribute to improving its own training process. Not a future concern anymore. They're saying it's close. And the fact that this warning is coming from inside the lab, not from external critics, changes the weight you should give it.

Priya: Welcome to AI Revolution for Tuesday, November 3rd, 2026. I'm Priya Nair.

Sam: And I'm Sam Kim.

Priya: We've got a packed show today. We're going to dig into Anthropic's RSI warning and what it actually means technically. Then we'll talk about the emerging black market for AI chip memory that U.S. export controls accidentally created. Google and Meta teamed up to open-source a PyTorch backend for TPUs. There's a fascinating mechanism design approach to AI alignment. Anthropic shipped a real latency improvement for long conversations. Goldman Sachs reversed course on giving AI coding tools to non-engineers. And we'll touch on OpenAI's latest hire from the Pentagon and some compute economics pressure. Let's get into it.

Sam: So RSI. Let's be precise about what this means, because the term gets thrown around loosely. Recursive self-improvement is the point where an AI system can do meaningful work on the pipeline that produces the next version of itself. That could mean improving training data curation, optimizing hyperparameters, writing better reward functions, even proposing architectural changes that actually pan out. The key word is "meaningfully" — we already have AI systems that assist with parts of ML engineering. The question is whether the system's contributions compound. Can the improved system then make even better improvements, creating a feedback loop?

Priya: And that's where the exponential acceleration concern comes from. It's not that the AI writes one good training script. It's that each cycle of improvement makes the next cycle faster and more effective.

Sam: Exactly. And what makes Anthropic's statement notable is the specificity. They're not saying "someday this could happen." They're saying they can see it from where they're standing. Now, there's a spectrum here. At one end, you have AI that's a really good ML engineering assistant — speeds things up, finds optimizations humans would miss. At the other end, you have the full recursive loop where human involvement becomes optional. Anthropic seems to be saying they're approaching the first part of that spectrum, and they're worried about the trajectory toward the second.

Priya: For practitioners, the operational question is: what does this change about your planning horizon? If capability jumps accelerate — even modestly — the assumptions baked into your security posture, your compliance frameworks, your model evaluation cadence, all of that needs to account for faster change. You can't evaluate models annually if the models are materially different quarterly.

Sam: And there's an honest uncertainty here. We don't know if RSI produces a smooth acceleration or a step function. The doomers worry about a discontinuity — a sudden jump. The optimists think it'll be gradual and manageable. Anthropic is basically saying: we should prepare for the former while hoping for the latter.

Priya: Let's shift to something very concrete and very messy. Huawei is now telling customers to go find their own high-bandwidth memory for AI chips. Sam, walk us through what's happening here.

Sam: So Huawei makes the Ascend series of AI accelerators, which are China's answer to NVIDIA's data center GPUs. These chips need HBM — high-bandwidth memory — the same stuff that goes into H100s and A100s. HBM is made primarily by SK Hynix, Samsung, and Micron, all of which are subject to U.S. export controls. Huawei can design the logic die, but they can't get the memory through legitimate channels. So they're telling customers: that's your problem now.

Priya: And a whole smuggling infrastructure has emerged to fill that gap. The specific loophole is interesting — smugglers bundle HBM chips with low-end processors that aren't export-controlled, ship the package through Hong Kong as a non-restricted product, then physically separate the HBM on the other side.

Sam: The price signal tells the story. HBM3E units are going for roughly four thousand dollars on this gray market, which is a massive premium. That price reflects both scarcity and risk. And here's the strategic takeaway: the export controls aren't denying access to the technology. They're taxing it. They're making Chinese AI compute more expensive and less reliable, but they're not preventing it. Whether that's sufficient depends on what you think the policy goal actually is.

Priya: Now for something that could reshape the compute landscape more constructively. Google and Meta jointly open-sourced TorchTPU — a native PyTorch backend for Google's TPU hardware.

Sam: This is a bigger deal than it might sound. PyTorch is the dominant framework for ML research and increasingly for production. But if you wanted to run PyTorch on TPUs, you had to use PyTorch/XLA, which required rethinking parts of your code to work with XLA's compilation model. It was functional but added friction. TorchTPU provides a native backend — it supports both eager execution, where operations run immediately like normal Python, and compiled execution for performance. It handles distributed training across TPU pods and exposes hardware-specific optimizations without requiring you to restructure your model code.

Priya: The competitive dynamics here are fascinating. Google has world-class training hardware that's been underutilized outside Google because of the software ecosystem lock-in. Meta has the largest PyTorch user base in the world. By collaborating on this, Google gets more customers for TPU Cloud, and Meta gets another viable option besides NVIDIA for its enormous compute needs. Both benefit from reducing NVIDIA's leverage.

Sam: If you're running large training jobs and your NVIDIA allocation is constrained or expensive, this just became a realistic alternative with much lower switching cost.

Priya: Let's talk about a research approach to AI alignment that's genuinely clever. MIT Technology Review covered work on applying mechanism design — sometimes called reverse game theory — to keeping AI agents in check.

Sam: So traditional game theory takes the rules of a game as given and asks: what's the optimal strategy? Mechanism design goes the other direction. You ask: what rules should I design so that the optimal strategy for every player is the behavior I actually want? Auction theory is the classic example — eBay's auction format is designed so your best move is to bid your true valuation. The researchers are applying this to AI alignment. Instead of trying to constrain an agent's behavior through rules it might learn to circumvent, you design the reward landscape so that following the desired behavior is genuinely optimal for the agent.

Priya: The analogy I find helpful: it's the difference between putting up a fence and making the grass greener on the side you want the cow to stay on.

Sam: Right. And this matters especially for multi-agent systems, where you have multiple AI agents interacting. Behavioral constraints get brittle fast in multi-agent settings because agents can find cooperative strategies to work around restrictions. Mechanism design is inherently about multi-party incentive alignment — it's built for that problem. It's still early-stage research, but as an architectural primitive for agentic governance, it's worth tracking.

Priya: Staying with Anthropic — they shipped Message Threads for the Messages API, and the performance numbers are meaningful.

Sam: The core problem: in a long multi-turn conversation, every API call traditionally retransmits the entire conversation history. At 500K to a million tokens, that's a lot of redundant data processing. Message Threads implements delta transmission — you send only what changed since the last turn. Time to first token drops about 35 percent for those long-context requests.

Priya: For anyone building agentic workflows where the agent maintains a long-running context — think coding assistants working through a large codebase, or research agents accumulating information over many steps — this directly reduces latency at the point where it hurts most.

Sam: And tying this to the compute economics story — Anthropic's margins on inference are reportedly getting worse, not better, even as they scale. Long-context requests are particularly expensive. This kind of optimization is partly a product improvement and partly Anthropic trying to make their own unit economics work. Those two motivations are aligned for now, which is good for users.

Priya: Quick hit on Goldman Sachs. They gave traders and bankers access to Claude Code and GitHub Copilot. Their CIO publicly praised non-technical staff doing vibe coding as recently as July. Now they've pulled access for everyone except engineers, citing productivity and risk management concerns.

Sam: This is an important data point. The gap between "this is impressive in a demo" and "this is safe to deploy in a regulated environment with non-technical users" is real and measurable. Goldman discovered that gap the hard way.

Priya: And briefly — OpenAI hired Dan Driscoll, former Secretary of the Army, as Senior Counselor for National Security Modernization. He's close to VP Vance, has bipartisan relationships in Congress, and this is clearly about positioning OpenAI for defense and intelligence contracts as those budgets take shape.

Sam: Looking ahead — the RSI conversation is the one that frames everything else. If Anthropic's assessment is accurate, we're entering a period where the rate of AI capability improvement could itself start accelerating. That changes planning horizons for everyone building on or governing these systems.

Priya: And I'd watch the TorchTPU story closely. If TPU adoption meaningfully increases among the PyTorch community, it shifts the compute market in ways that affect pricing, availability, and the competitive dynamics between cloud providers. That has downstream effects on who can afford to train frontier models and where.

Sam: The mechanism design work is one I want to see develop. We need alignment approaches that scale with capability, not against it. Incentive design is a more robust foundation than behavioral restriction. The research is early, but the direction is promising.

Priya: That's our show for Tuesday, November 3rd. Show notes and links to every story we covered are at cleartext.fm.

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