For three weeks the sovereignty question climbed the stack — the model, the cost, the rails. This week it changed shape: who checks the work? The verification bottleneck appeared in four unrelated domains at once — peer review, code review, public benchmarks, and science's publication commons — while trust migrated from shared public instruments to private ones.
OpenAI published 722 mathematical manuscripts from an unreleased internal model — claimed solutions to 90 of the top 500 open problems, including the Quasi-Riemann Hypothesis, some in roughly three hours of compute each. The output is public; the capability that made it isn't. Peer review runs on months per paper; the proofs arrived in an afternoon. By one tally, that afternoon potentially contains 81% of the major maths discoveries of the last three years. Steve Hsu's proposal: mathematics splits into a vast machine mathematics mostly consumed by AIs, and a compressed human mathematics — an "effective theory" of the machine frontier. Many results are Lean-verified — provable but incomprehensible. Trust migrates from understanding to instrument, and the frontier bifurcates away from public access for the second week running.
The week's cleanest convergence — at least four independent sources, same direction. Reflection AI shipped Beam: 501B total / 23B active parameters, Apache 2.0, US-trained from scratch, claiming 80.9 on SWE-bench Verified. NVIDIA completed a $12.9B acquisition of Hugging Face. Commerce Secretary Lutnick considered directly funding US open models. Palantir's Karp: customers "want to own the means of production." Microsoft's Nadella published the "Reverse Information Paradox" — with closed models you pay for intelligence twice, once in money and once in the proprietary knowledge you reveal. The demand side has moved: 60% of Vercel AI Gateway spend and a record 78.4% of tokens on a single day went to open weights. Microsoft's counter-play — "make our model your model" — imitates the challenger's selling points. And the price war continued: Claude Haiku 5.5 matches GPT-6 Luna pricing at ~75% cheaper than its own predecessor. Weights you can own, fine-tune, and run where you choose — the difference between owning a tool and renting a dependency. And you can't check a model you can't run.
Ethan Mollick published a rare public mea culpa: he was wrong that humans would manage agent teams like managers run companies — swarm approaches beat careful human orchestration. Another win for the Bitter Lesson. Same week, Gergely Orosz's LDX3 keynote: "practically nobody writes code by hand anymore," the IDE is fading, and "code reviews became theatrical" — the ceremony remains, the function has drained away. Quality and reliability are down. Meanwhile public benchmarks are functionally dead — they're in every training set, and the only benchmark that matters is one built on your own tasks. Private evals become the moat. The oversight layer is becoming performance across every domain. The infrastructure answer arriving: Stacklok — Kubernetes' co-creators, $17.5M Series A — betting that control loops, not code review, are the replacement oversight technology: stop supervising outputs, instrument the runtime.
Three domains moved in one week, in sequence. Mathematics: the 722-paper dump. Materials: Periodic Labs — founded by ex-OpenAI's Liam Fedus and ex-DeepMind's Ekin Dogus Cubuk — applying the frontier-lab scaling playbook to matter itself, "from semiconductors to superconductors." Cells: the Chan Zuckerberg Biohub's Virtual Biology Initiative grew from $500M to $1.8B — NIH sharing $500M+ of federally-funded datasets, DOE committing $500M+ over five years, DeepMind/Meta/Isomorphic adding $300M, with a one-year private data embargo before release. Zuckerberg is explicitly framing biology as "next after math's AI moment." The quiet logic of the sequence: language models argue with you about whether they're right — materials and cells don't. When human checking can't scale, the proving ground moves to where reality is the referee.
Three threads from previous weeks continued underneath: the interface war (GPT-6 off the paywall, ChatGPT rebuilt as a visual canvas, Claude embedded in Google Workspace), the cost collapse (Haiku 5.5 at Luna pricing, GPT-6 free), and the governance convergence — David Robinson, who drafted OpenAI's Preparedness Framework, quit calling the culture broken and arguing labs need nuclear-plant-style redundancy, the same week the "safe acceleration" doctrine crystallised in Washington. Opposite camps, same conclusion: the current culture can't hold the capability.
The new pattern is the bottleneck moving from generation to verification. Peer review can't absorb the proofs. Code review can't absorb the code. Public benchmarks can't rank the models. Science's publication commons is being replaced by negotiated access. The old control strategy — review the outputs — is dying faster than the experts predicted. The replacement isn't checking more; it's different instruments: evals built on your own tasks, weights you can run where you choose, control loops that instrument the process.
The public-good question is sharp this week. Left alone, trust migrating to private instruments ends with verification for those who can afford it and vibes for everyone else. But the best verification tooling — Lean, open eval harnesses, open weights — is open. The commons can scale with the generation layer, if we choose to build it that way. You can't check 722 proofs. Nobody can. But you can build the toolchain that checks the class of thing you depend on — and refuse systems that won't let you.