The Bank Problem, Zero Credentials, and 722 Math Proofs

00:00:04 OpenAI published a collection of 722 mathematical manuscripts across 372 families of related results. The model that produced them is unreleased — you can't run it, only read its output. The collection came from an evaluation of about 4,000 research problems. Each one took roughly three hours of ChatGPT Pro compute on average.

00:00:28 Some papers include Lean proofs; others remain unverified. The headline result is labeled Result 003 — the Quasi-Riemann Hypothesis. According to coverage from Smol AI, it's somewhere between a Fields Medal result and what one mathematician called "the biggest result in number theory in 200 years." There are also results touching integer multiplication algorithms, an elastic inverse problem open since 1994, and partial progress on Hodge and Birch–Swinnerton-Dyer.

00:01:03 About twenty percent of the collection consists of disproofs or counterexamples. An analysis at Latent Space noted that this chunk undercuts earlier descriptions of these results as pure brute-force search victories. OpenAI consulted the Institute for Advanced Study's independent Advisory Group on Mathematics and AI before releasing them.

00:01:28 The model itself stays locked away. What you're looking at is a public artifact repo, not a model card. The archive just has text files on GitHub ranging from number theory to elastic inverse problems — the real test is whether any of them survive human scrutiny long enough to matter.

00:01:49 Which proofs mathematicians cite first will tell us whether this collection shifts anything.

00:01:56 Rep. Lori Trahan is rolling out legislation to clarify developer liability when AI agents cause harm, framed against the rogue agent hack of Hugging Face earlier this year. Nicholas Wu reported the story on Wednesday. Miles Brundage, who works on AI safety at Scale AI, amplified it.

00:02:17 Trahan's office centers on direct causation — agent action, agent harm, or developer responsibility — rather than a sweeping "AI caused something bad" category that most legislation has been circling around. The proposal tries to draw a line between what platforms ingest from untrusted agent outputs and what crosses into actionable developer liability.

00:02:42 That boundary is where every AI company with API access writes its terms of service, and it's also where error handling actually gets tested outside benchmark evaluations. It shows up when an agent follows its objectives through to something that breaks a bank account or leaks data across a corporate boundary.

00:03:05 Trahan has chaired the Congressional AI Caucus and focused on platform accountability before this. How they define "direct response" versus "foreseeable downstream effect," and who draws the line when a chain of agent actions compounds, will set the baseline for developer terms over the next few years.

00:03:26 Platforms that write their definitions first will shape the rest.

00:03:31 Docker's Jim Clark gave a talk on Wednesday proposing MCP gateways for zero-credential agent sandboxes. The core idea is simple, though the engineering implications are less obvious than they sound at first pass. Every MCP call from an agent routes through a single gateway endpoint inside each sandbox.

00:03:53 That endpoint controls which tools, resources, and prompts are available to the harness at any given time. You can swap models — Claude, Codex, Gemini — without reconfiguring MCPs for each one. Clark's working rule is straightforward: how many credentials should live in a sandbox?

00:04:13 The answer is always zero. Keeping tokens out of the image or environment means damage from a compromise stays near nothing. To handle real authentication, Clark and partners at Entropic and Octo are building XAA — Cross-App Application — which lets agents exchange identity claims with existing corporate SSO providers.

00:04:37 An agent picks up an ID JWT and authorization grant through your current OAuth infrastructure without forcing new consent flows. The talk used a newsroom analogy to map it out: a researcher sandbox reads the web but can't write to anything. A fact-checker sandbox has no network access but can pull filtered research.

00:05:00 A reporter sandbox publishes through one MCP with zero external reach. Each individual agent carries far less capability than the workflow as a whole, and that division of labor is where the security model actually holds. Capability needs to match task intent rather than the full inventory of what exists on the machine.

00:05:23 If you're committing code during an hour-long coding session, your sandbox needs Git signing keys for two minutes, not the whole hour. Docker's sbx CLI already implements the containerized version of this approach. It's early engineering on a problem most builders will run into within months.

00:05:44 Two separate voices made points that land on the same structural question. Naval posted that models are the last moat in software, then expanded the argument: AI can train on essays and rewrite them, decompile and recode software, ingest art and recreate it. But he suggests AI itself doesn't want to be distilled.

00:06:07 More software will retreat to the server to resist extraction. If the moat shifts from data to inference, the winners are companies that own fast, cheap compute and distribution — not those holding the best datasets or downloadable models. Dor Sasson, co-founder and CEO of Stigg, arrived at a related conclusion from an architectural angle.

00:06:31 He told an audience on Wednesday that every AI company is accidentally building a bank. His framing traced back to April 2024, when Anthropic cut third-party agent access to Claude Max subscriptions because it couldn't separate subscription consumers from API consumers.

00:06:51 OpenAI adjusted pricing five-fold overnight. GitHub froze and eliminated Copilot free trials. Those looked like pricing emergencies at the time. Sasson argued they were infrastructure failures — entitlement checks happening after the invoice rather than before, when a company could still stop the drawdown.

00:07:12 He pointed to OpenAI's own financial engineering team publishing a decision-waterfall architecture with synchronous runtime checks, hold-and-settle concurrency handling, double-entry bookkeeping for credit pools, and multi-source drawdown priority. That is bank-grade accounting applied directly to AI consumption.

00:07:35 Naval frames the moat around resistance to extraction; Sasson frames it around the infrastructure layer that decides who consumes what and when. Both are tracking the same shift in how AI commerce actually survives commoditization.

00:07:52 Which of OpenAI's 722 manuscripts get cited by human mathematicians first will set the baseline for how this collection is read over the next few years. Twenty percent of that archive consists of disproofs and counterexamples, which means the proofs need scrutiny.

00:08:08 Any single proof surviving two years of independent verification carries weight, regardless of where you land on AI-generated math. How platforms respond to Trahan's proposal will define developer liability for this decade of agentic systems. The direct-causation language in agent-triggered harm is where the rules get written for the first time, and the platforms that draft their definitions earliest will shape the rest.

00:08:33 Which proofs survive first will tell us whether this collection shifts anything. Seln Oriax.