Sam: So over the weekend, Anthropic's Claude Mythos solved a math conjecture that had been open since 1946. The Erdős unit-distance conjecture. And what makes this remarkable isn't just that it solved it — it's that OpenAI had separately disproved the same conjecture just days earlier, and Mythos apparently came at it from a completely different angle and produced what Anthropic engineer Sholto Douglas called a "cute, simple proof." Two frontier models, independently operating at the level of top-tier mathematical research, on the same problem, in the same week. That's where we are.

Priya: Welcome to AI Revolution for Wednesday, May 27th, 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 what the Erdős result actually tells us about frontier model reasoning. Then we're covering a critical vulnerability affecting millions of AI agent deployments, Nvidia's staggering Taiwan supply chain numbers, China's AI-powered surveillance upgrade, Robinhood letting AI agents trade real money, Google's AI Search backlash, and a startup using India's gig workers to collect physical training data for robots.

Sam: Let's start with the math. The Erdős unit-distance conjecture, posed by Paul Erdős in 1946, asks a deceptively simple question: if you have n points in the plane, how many pairs of those points can be exactly distance one apart? Erdős conjectured the answer was at most n to the power of one plus epsilon — basically slightly superlinear — for any epsilon greater than zero. Generations of mathematicians have chipped away at this. The best known upper bounds have used techniques from incidence geometry, and the problem has remained stubbornly open.

Priya: And what's interesting is the convergence. OpenAI disproved the conjecture — meaning they found a construction that beats the conjectured bound — and then Mythos apparently produced a proof approaching the problem from a different direction. Can you unpack what "cute, simple proof" might mean here?

Sam: Yeah, so in mathematics, a "cute" proof usually means it avoids heavy machinery. It finds an elegant path that makes you say "oh, of course" after the fact but that nobody saw beforehand. The significance here is twofold. First, the result itself matters to combinatorial geometry and has downstream connections to computational geometry problems that show up in algorithm design. But second, and maybe more importantly for our audience, what Douglas is calling "serious overhang" — his phrase — suggests that these models have latent mathematical reasoning capability that hasn't been fully tapped yet. They're not just pattern-matching known proof strategies. They're finding novel arguments.

Priya: The cryptography implications are worth flagging here. If frontier models can find novel proofs in combinatorial number theory, the question of when they start making progress on problems closer to the foundations of cryptographic hardness assumptions gets more urgent. We're not there yet. But the trajectory is clear, and it's faster than most people in the field expected even six months ago.

Sam: Agreed. And the fact that two different labs, two different models, converged on the same problem within days suggests this isn't a fluke of one model's training data. There's something about how these architectures reason about mathematical structure that's genuinely improving.

Priya: Let's shift to something that's immediately actionable. A critical vulnerability called BadHost has been discovered in Starlette, the Python async web framework that sits underneath a huge portion of the AI agent stack. Starlette gets about 325 million weekly downloads. If you're running FastAPI — which is built on Starlette — or really any modern Python-based LLM serving infrastructure, you're probably affected.

Sam: Right. Starlette handles the HTTP request layer for an enormous number of AI agent deployments. It's one of those invisible foundational dependencies. FastAPI, which is the go-to framework for standing up model endpoints, is essentially a layer on top of Starlette. So when there's a vulnerability in Starlette's request handling, it propagates through the entire stack.

Priya: The details on the exploit mechanism are still being characterized, but the name "BadHost" suggests it's related to host header handling — potentially enabling host header injection, which can lead to cache poisoning, password reset hijacking, or in the worst case, server-side request forgery. In agent architectures where you have services calling other services, where agents are making HTTP requests on behalf of users, a host header vulnerability can be particularly dangerous because the blast radius scales with the agent's permissions.

Sam: And this highlights something we keep coming back to: the agentic AI stack is being built on the same open source infrastructure that everything else runs on, but the security audit attention hasn't kept pace with the deployment velocity. Starlette is well-maintained, but 325 million weekly downloads means an enormous attack surface. If you're running any production agent system, check your Starlette version today.

Priya: Moving to infrastructure. Nvidia's yearly spending in Taiwan has gone from $15 billion to $150 billion. A tenfold increase, concentrated almost entirely at TSMC.

Sam: That number is just enormous. To put it in perspective, $150 billion annually at a single country's suppliers — that's larger than the GDP of many countries. And it's essentially all going to one thing: fabricating AI accelerators. The demand from hyperscalers for H100s, B200s, and whatever comes next has created this incredible concentration of spending.

Priya: The technical story here is that advanced AI chips require TSMC's leading-edge nodes. There is no realistic alternative supplier at the 3-nanometer and below process nodes that Nvidia needs for its latest architectures. Intel Foundry and Samsung are years behind on yield and volume. So when we talk about geopolitical risk in AI infrastructure, this is what it looks like concretely: the entire AI compute buildout runs through a single island in the Taiwan Strait.

Sam: And it's a single point of failure that you can't engineer around quickly. Building a new leading-edge fab takes three to five years and tens of billions of dollars. The CHIPS Act facilities in Arizona and elsewhere are coming, but they won't materially change this concentration before 2028 at the earliest.

Priya: Now, China's surveillance story. Manufacturers like Hikvision and Huawei are shipping cameras with embedded computer vision and language models that run on-device. Officers can type natural language queries — "show me anyone who entered this intersection carrying a large bag between 2 and 4 AM" — and the system returns results across a distributed camera network.

Sam: What's technically notable here is the edge deployment. Running multimodal models — vision plus language — on camera hardware means these systems don't need to stream all video to a central server for processing. The inference happens locally, which dramatically reduces bandwidth requirements and makes the system work even with China's existing network infrastructure. It also means the cameras can autonomously flag behaviors in real time without waiting for a round trip to the cloud.

Priya: Human Rights Watch has flagged the civil liberties implications, and they're severe. But from a technical architecture standpoint, this is also a proof of concept for what edge AI deployment looks like at genuine national scale. Millions of devices, heterogeneous hardware, running inference continuously. The engineering challenges they've solved around model compression, power efficiency, and distributed querying are significant regardless of how you feel about the application.

Sam: And the natural language query interface is the part that changes the operational picture. Previously, surveillance cameras generated footage that humans had to review. The bottleneck was always human attention. Now the bottleneck is removed. You can search behavioral patterns across an entire city's camera network the way you'd search a database.

Priya: Shifting to financial markets. Robinhood is now letting users create a dedicated sub-account, fund it with a specific balance, and hand control to an AI agent that can buy and sell stocks autonomously. No per-trade human approval.

Sam: This is a genuinely novel deployment pattern. The agent gets a sandbox with real money and real market access. It's not paper trading. The architectural choice to use a separate sub-account with a defined balance is interesting — it's essentially a blast radius limiter. You're saying: this agent can lose at most X dollars.

Priya: But the regulatory questions are fascinating. Who's liable when an agent executes a trade? Existing securities law assumes a human decision-maker. Market manipulation rules are written around concepts like intent. Does an AI agent have intent? If a thousand users deploy similar agents and they all pile into the same stock simultaneously, is that coordinated action? These are open legal questions, and Robinhood is essentially forcing regulators to address them by shipping the feature first.

Sam: Quick hits now. DuckDuckGo app installs spiked 30% after Google replaced traditional search results with AI agents as the default experience at I/O 2026. This is a real signal. A meaningful segment of users actively doesn't want AI-mediated search. They want links, they want to evaluate sources themselves. Google is betting that AI-first search is the future, and they may be right long-term, but the transition cost in user trust is measurable and immediate.

Priya: And then there's Human Archive, a startup out of UC Berkeley and Stanford. They're paying gig workers in India to wear camera-equipped caps and multi-sensor devices while performing everyday physical tasks. The goal is collecting embodied interaction data — how humans grasp objects, navigate spaces, react to physical environments — because this data simply doesn't exist on the internet. You can scrape text and images at scale, but you cannot scrape what it feels like to pick up a coffee cup or walk through a crowded market.

Sam: This is addressing the core data bottleneck for physical AI and robotics foundation models. Simulation helps, but sim-to-real transfer is still a massive challenge. Real-world sensor data from diverse environments — different lighting, different objects, different cultural contexts for how people move through spaces — is incredibly valuable. India's gig economy gives them scale and environmental diversity at a cost point that makes the data collection economically viable.

Priya: One more story worth noting. IEEE Spectrum published an analysis of South Africa's draft AI policy. South Africa controls about 88% of global platinum-group metal reserves, which are used in components across the semiconductor and data center supply chains. They also host the largest data center market in Africa. Both the US and China are actively competing for AI infrastructure positioning there. But the draft policy doesn't leverage any of this.

Sam: It's an interesting case study in how resource-rich countries are thinking — or not thinking — about their position in the AI supply chain. Platinum-group metals aren't as central to chip fabrication as, say, TSMC's process technology. But they matter for specific components, and the geopolitical positioning matters a lot.

Priya: Looking ahead, Sam. What are you watching after today?

Sam: The math discovery story is the one I keep coming back to. If we're seeing serious overhang — if these models have more mathematical reasoning capability than we've extracted — then the question is what happens when labs start systematically pointing them at open problems. Not just as demos, but as a sustained research program. The rate of mathematical discovery could accelerate in a way we haven't seen before, and that has implications for fields well beyond pure math. Optimization, physics, cryptography — anything with deep mathematical structure.

Priya: And I'm watching the infrastructure fragility angle. The Starlette vulnerability and the Nvidia-Taiwan concentration are two different flavors of the same underlying issue: the AI stack is scaling faster than its foundations are being hardened. Whether that's software supply chain security or geopolitical supply chain diversification, the gap between deployment velocity and infrastructure resilience is widening. That's going to produce more incidents and more surprises.

Sam: Well said. That's the tension of this moment — extraordinary capability advancement on one axis, and fragility accumulation on the other.

Priya: That's our show for Wednesday, May 27th. Show notes and links to all the stories we covered are at cleartext.fm.

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