A packed Innermost Loop day: OpenAI’s math-research publicity collided with community backlash, fresh model launches (GPT-6, Claude Haiku 5.5), and the usual commercial scoreboard (revenue run rate, API price war). On the open-weight side, Mistral Large 4 / Le Chonk kept trending while Ecosia publicly swapped off Mistral for Qwen-class models.
Today
- Welcome to October 9, 2026 — Alexander Wissner-Gross’s daily Innermost Loop edition (math drama, model launches, policy). The Innermost Loop
- Sharing AI progress in mathematics — OpenAI’s post on formal-math progress and Lean verification (center of today’s researcher controversy). OpenAI
- OpenAI is pissing off a bunch of mathematicians again — Wired on proof claims, community norms, and trust in public research announcements. Wired
- Erdős Problems forum froze proof claims — Moderation response to the flood of AI-assisted “solutions” on open problem threads. Erdős Problems
- Researchers report progress refuting 3SUM and APSP hypotheses — arXiv preprint circulating in the same news cycle (verify independently). arXiv
- GPT-6 for everyone — OpenAI announcement framing wider availability of its next flagship tier. OpenAI
- Claude Haiku 5.5 — Anthropic ships a faster Haiku generation for high-volume workloads. Anthropic
- The AI price war is heating up — OpenAI gaining on Anthropic — WSJ on API pricing pressure and share shifts between frontier vendors. Wall Street Journal
- OpenAI revenue run rate and hyperscaler ties — CNBC on ~$50B narrative and NVIDIA/Oracle/CoreWeave economics behind the stack. CNBC
- Impactful scheduling for GPU clusters — Allen Institute post on allocation policies that actually move utilization (HF blog). Hugging Face Blog
- Into the Omniverse: developers, simulations, and frontier AI agents — NVIDIA on combining Omniverse libraries with agentic dev workflows. NVIDIA Blog
- EmbeddingGemma 2 — Google’s updated open embedding model for developers building retrieval stacks. Google
- AWS, Upstage, and Ollama align on a decision-model API — Interop push for “decision” models (OpenAI not signed on). The New Stack
- Ecosia switches from Mistral to Qwen, GLM, and Kimi — European search operator cites quality, capacity, and sovereignty. TechNode
- Mistral Large 4 trails some Chinese open models in independent tests — Tom’s Hardware benchmark sanity check on the trillion-parameter preview. Tom’s Hardware
- Artex AI agent goes closed-source after South Korean bank hack — Developer response to misuse of pentest-oriented tooling in live attacks. The Straits Times
- Secret dates in system prompts undermine LLM evaluation — Benchmark contamination angle when eval harnesses leak calendar hints. Unite.AI
- Introducing Beam — Reflection.ai product/blog drop picked up in the AWG firehose. Reflection.ai
Watch: OpenAI
- OpenAI coding-agent spend (data note) — Epoch.ai estimate of daily burn on coding agents — useful for token-economics conversations. Epoch.ai
- Mathematicians urged boycott of OpenAI — Professional society statement in the backlash thread. ahmath.org
Watch: Claude / Anthropic
- 2026 usage policy update — Anthropic refreshes acceptable-use rules as enterprise adoption grows. Anthropic
- Anthropic subscriptions “offer 5×” (analysis) — SemiAnalysis on unit economics vs OpenAI pricing moves. SemiAnalysis
Watch: Open-weight models
- Mistral’s Large 4 “Le Chonk” joins the West’s open-model push — Launch narrative in the shadow of Chinese OSS leaders. The Rundown AI
- The agent alone is a myth — production is the harness — Syndicated Viren Baraiya quote aligned with operable-agent consulting thesis. BigGo Finance
Built with utilities/ai_report_build.py (24-hour lookback). Edited for publication. About the AI Report.