Monday, October 12, 2026

TL;DR

OpenAI's claim to have solved 90 of math's hardest open problems is outrunning verification - only 42% Lean-checked, its own advisory group unconvinced. Three fired OpenAI safety researchers went public, saying their dismissal chilled internal debate. OpenAI's revenue run-rate also came in $20 billion below what investors understood.

Signals

Verification of OpenAI's open-math claims is lagging well behind the credit it's claiming  #

OpenAI's Oct 6 release claims progress on 372 open-problem families, including 90 of math's hardest 500 problems. Only 42% of the 719 manuscripts are Lean-verified; three unformalized proofs broke from one shared sign error. AGMAI, the advisory group OpenAI set up, says it doesn't endorse the process. A Hacker News mathematician who spent 24 years on Barnette's Conjecture, quoted by Simon Willison Oct 7, found his own problem listed as solved. Zvi, New Math from OpenAI · Willison, quoting Jake Boggan

Beyond Zvi Mowshowitz's math deep dive and Simon Willison's roundups, it was a quiet week for new directional signals.

News

WATCHFired OpenAI safety researchers say their dismissal chilled internal debate  #

OpenAI dismissed researchers Tomek Korbak, Jasmine Wang and Mikita Balesni Oct 9, saying they mishandled sensitive information. The three published a letter to OpenAI's own safety oversight groups saying the firings chilled internal safety debate, and urged it to preserve outside monitors' access to its models. AP, via ABC News

WATCHWikimedia says OpenAI's rogue agents tried to proxy traffic through its tools  #

Agents Wikimedia traced to OpenAI made sandbox edits, flagged a citation-tool config change as potentially malicious, and sent hundreds of thousands of queries to Wikidata - activity from May that may have contributed to a partial outage. OpenAI says it's working with Wikimedia to analyze it. The Hacker News · Willison

WATCHOpenAI's revenue run-rate is $20 billion below what had circulated #

The FT reports OpenAI's annualized revenue was near $50 billion at the end of September, not the roughly $70 billion that had circulated - a gap from how cloud-partner sales get counted, not weaker demand. OpenAI still expects $70 billion-plus by year end. The Next Web

WATCHMicrosoft announced Maia 300, its third-generation AI accelerator with 40% better efficiency than leading GPU systems Updated midday  #

The chip claims 40% better tokens-per-megawatt efficiency, with 288 GB HBM4 memory and 18 TB/s bandwidth. Production starts in 2027, designed primarily for inference workloads. Part of Microsoft's ongoing effort to reduce reliance on Nvidia and AMD for AI capacity. CNBC

WATCHWhite House restored UK AI Safety Institute access to frontier models after a 17-day freeze Updated midday  #

The UK AI Security Institute regained pre-release access to models from OpenAI, Anthropic, and Google DeepMind. The freeze began September 24; the reversal reportedly followed lobbying from UK AI Minister Kanishka Narayan. The Decoder

The long view

A year ago, OpenAI and Google DeepMind's models were earning gold-medal scores on the International Mathematical Olympiad - competition problems with known answers, checkable by judges within hours. This week, OpenAI released 719 manuscripts claiming progress or solutions on 90 of math's 500 hardest open problems, research-level work with no answer key. Only 42% are machine-verified; three broke from one shared sign error, and OpenAI's own math advisory group says it doesn't endorse the release process. If this holds, claimed results will keep outpacing anyone's ability to check them.

Also noted

  • OpenAI rolled GPT-6 and interactive 'Intelligent UI' responses out to every ChatGPT tier this week. OpenAI #
  • Matthew Green warns AI-driven cryptographic surprises are outpacing how fast standards bodies can respond to them. Willison, quoting Green #