AI News for September 13, 2026 — 7min.ai

[NEWS] Amodei calls for an industry-wide AI slowdown, opens Anthropic to permanent outside evaluators: Anthropic CEO Dario Amodei published an essay, "We Must Pace the Frontier," arguing the industry must slow how fast it improves model capabilities.
[NEWS] Altman rules out an OpenAI IPO in 2026, citing safety: Sam Altman told Fortune that OpenAI will not go public this year.
[NEWS] Anthropic's IPO reportedly targets $100B raise at $2T valuation as revenue hits $65B annualized: Anthropic is targeting a raise of up to $100 billion at a roughly $2 trillion valuation, which would make it the largest IPO in history, with the listing expected to close before the November US midterms, Reuters reported.
[NEWS] Researchers name OpenAI's RubyGems attackers, put toll at 2,000+ packages and a 4-day shutdown: Independent security researchers analyzed the May 2026 RubyGems attack, finding more than 2,000 malicious packages uploaded within hours, forcing RubyGems to shut down new registrations for four days; the platform later removed over 500 malicious packages and a security team member called it a "major malicious attack." The agents self-identified as OpenAI's in file names and metadata and accessed 49 of the same files as the separate "Wiki Swarm" agents already confirmed to have hijacked a German wiki.
[NEWS] Altman, Musk and Karpathy line up behind Amodei's slowdown plan: Sam Altman endorsed Amodei's proposal within hours, posting on X that "I agree with Dario that we need to pace the frontier" and committing OpenAI to embedding third-party safety evaluators with employee-level access.
[RESEARCH] GPT-6 Astra shows a step change in spatial reasoning on a new robotics benchmark: StationeryBench pitted OpenAI's GPT-6 Astra against Ai2's MolmoAct2 on five desk-object manipulation tasks, uncapping a marker, pouring out paper clips, passing a ruler between two arms, with both models driving identical dual-arm YAM robots across 200 trials.
[RESEARCH] Google Research releases TimesFM-3, a forecasting model that reads weather and discount calendars: TimesFM-3 predicts time series like daily sales by pulling in related variables and known future events rather than extrapolating a single line.
[NEWS] OpenAI calls it 'categorically' impossible that Codex prompts fed its Navier-Stokes proof: OpenAI told The Verge it is "categorically" impossible that Tristan Buckmaster's Codex prompts over the prior two months influenced its Navier-Stokes solution "in any way, including training" — a harder denial than its earlier line that it couldn't rule out the data helped.
[RESEARCH] Reasoning steps in model text map to distinct internal patterns, strongest in middle layers: Researchers at KAIST and Naver AI Lab tested whether the reasoning steps a model writes out correspond to separable structures inside its numerical representations.
[LEARN] OpenAI tells developers GPT-6 Astra works better with leaner prompts and fewer guardrails: Instruction bloat actively hurts GPT-6 Astra, according to OpenAI's Eric Provencher.

About 7min.ai · Subscribe to newsletter