Claude lifts Riemann zeta zero bound to 67.2 percent
Anthropic published the most striking AI-math result in weeks, while a Hugging Face–EleutherAI project found a cheap fix for one of open training data's quietest problems — and Hong Kong's biggest broadcaster decided it wants in on the compute buildout.
An unreleased research version of Claude improved a long-standing lower bound on the proportion of Riemann zeta zeros that lie on the critical line — from 41.6 percent to 67.2 percent — and produced a formally verifiable proof of the result. The work began when Jarred Sumner, a non-mathematician on Anthropic's staff, prompted Claude to "take a real stab" at the Riemann hypothesis. After generating 650 ideas that went nowhere, Claude spent a day and a half coordinating about 60 subagents that ran 2,400 shell commands, wrote hundreds of Python scripts, checked its finding against 54 arXiv papers, and refereed one another's work — 31 million output tokens in all. Claude wrote the result up as a paper itself and recommended that human mathematicians validate it; two Anthropic mathematicians did, and external experts Brian Conrey and Dan Goldston reviewed it on short notice.
Anthropic is careful to say the techniques probably won't crack the Riemann hypothesis itself, a 167-year-old problem carrying a million-dollar bounty. The signal is the process: an AI that formulates a research program, executes it with subagents, checks its own work, and insists on human verification — the same shape as the ChatGPT-assisted proof that toppled the HRT conjecture this weekend, which we covered — ChatGPT-assisted proof topples the 30-year-old HRT conjecture. The question is no longer whether models can do novel math; it's how much of the scientific workflow they can drive on their own.
Hugging Face and EleutherAI's FineBooks project benchmarked 14 open-weight OCR models on 2,165 historical book pages — and found small models beat big ones at cleaning up the text AI trains on. The top performer, a 3-billion-parameter model called dots.mocr, hit 97.6 percent character accuracy at $1.94 per thousand pages; the 0.9-billion-parameter OvisOCR2 was nearly as accurate at 46 cents, while a model three times dots.mocr's size scored lower. The stakes are real: a prior project found a model trained on old OCR text learned at only 30 percent the efficiency of one trained on human transcripts, and Common Pile's roughly 300,000 public-domain books are full of legacy OCR errors. FineBooks plans to reprocess about 200,000 documents from the Biodiversity Heritage Library and release the cleaned text as an open dataset — an unusually cheap upgrade to the open AI commons, though the team says accuracy still falls short of scholarly standards.
Hong Kong broadcaster TVB plans to enter the AI compute business, announcing a joint venture with real-estate investor Gaw Capital to build a GPU facility at its Tseung Kwan O campus. TVB would hold 51 percent and Gaw Capital 49 percent, with Gaw committing up to HK$2 billion (about $256 million) in equity; the first phase, aimed at roughly 10,000 petaFLOPS, is targeted to start operating in the fourth quarter of 2027. The non-binding agreement also includes a letter of intent signed last month with "a leading global tech group" interested in buying compute capacity. It's a reminder that the AI buildout keeps pulling in unlikely players — a broadcaster betting that AI-generated content and external customers can justify a data center — though the deal still needs formal agreements and government approval.
What to watch: whether TVB's mystery global tech customer signs, and whether Anthropic ever ships the research version of Claude behind the zeta result.
If AI keeps producing novel math results, do we need new norms for crediting machine co-authors? Tell us in the comments.
Sources: Anthropic · Hacker News discussion · The Decoder · FineBooks dataset (Hugging Face) · Sing Tao Headline · The Paper · CLS
