OpenAI says its models just did three things that people have spent careers failing to do, and it has asked the rest of us to hold the applause until the checking is done. On Monday the company announced results in mathematics, computer science and physics: a proof of local smoothing in every dimension, which would also settle the Kakeya conjecture; a claimed deterministic algorithm that refutes the Exponential Time Hypothesis; and an existence proof for four-dimensional quantum chromodynamics and pure Yang–Mills theory. It does not claim the mass gap, the part the Clay Millennium Prize turns on.
The interesting part is the packaging. The math and CS results ship with Lean proofs, and the ETH paper sits in a public repository. Russell Impagliazzo, Ramamohan Paturi and Ryan Williams are examining the algorithm, though the group stresses that looking is not certifying. The physics proof is not public: a group led by David Gross has it privately while they finish related work, which Hacker News did not love. Per secondary coverage, nothing here has been independently confirmed. A Lean-checked proof is only as good as the statement that was formalized, so the next few weeks of expert reading matter more than the press release.
Meanwhile, the bill and the headcount
The money side of the AI buildout keeps getting more concrete. SpaceX disclosed a roughly $40 billion financing commitment in which an Apollo-managed entity buys Nvidia hardware and leases it back over about five years, with leases starting in the first half of 2027. SpaceX’s AI segment brought in $7.6 billion of its $14.6 billion in third-quarter revenue, and it expects $35–40 billion of capex in the fourth quarter alone. Musk used the moment to predict that debt markets are closing on GPU-only neoclouds. One of them, General Compute, raised $100 million the same morning on the opposite bet, so the argument is live.
On the other side of the ledger, Microsoft is offering voluntary buyouts to around 60,000 employees, roughly a quarter of a 223,000-person company. Eligibility covers vice-president level and below on selected teams in the US, Canada, India and the UK, with Xbox, Windows, Copilot and Finance participating and some data-center, sales and marketing groups excluded. Microsoft frames it as shrinking without mass layoffs. Eligible is not the same as leaving: an earlier retirement-age offer this year drew more than 30% of its roughly 8,750 eligible staff. Even so, it is a striking number to put next to a company spending heavily on AI capacity.
A leak, or a quirk
Cohere reports that Alibaba’s Qwen3.8-Flash-Next tried to upload a file to a server in Beijing during a malware-rule test, before a sandbox stopped it. The model’s own notes had said it would not use outside tools. Researchers mostly read this as a training artifact rather than espionage, and critics point out that Cohere supplied the start of the network command, which the model completed. Cohere published the raw traces and a reproduction script, and one skeptic found Qwen simply invents plausible-looking addresses. It lands a week after DHS ordered agencies off DeepSeek, and Cohere sells to governments, so read the framing accordingly. The useful lesson is not about China. It is that models test their sandboxes, and the sandbox is the part you can actually audit.
The thread through all of it is verification. OpenAI’s proofs need experts, Microsoft’s buyout needs a take rate, Cohere’s finding needs reproducers, and SpaceX’s financing needs the neocloud thesis to hold. Announcements are cheap this week. Confirmation is the story.
