The decision-model stack kept moving: Microsoft’s Decision-1 (Qwen-based scorers) and guides to OpenAI’s Decisions API. Deal rumors put NVIDIA closer to Reflection AI. Trust and safety dominated Anthropic headlines — a false homicide tip to Philadelphia police and new research on unintended model actions.
Today
- Welcome to October 10, 2026 — Innermost Loop daily edition (deal rumors, safety incidents, infra demand). The Innermost Loop
- Microsoft Decision-1: 9B Qwen model launch explained — shattered.io walkthrough of the scorer-style model Microsoft pitched for structured choices. shattered.io
- Microsoft’s Decision-1 scores options instead of writing text — Coverage claims large speedups vs GPT-class chat models for decision workloads (treat benchmarks as vendor-adjacent). XenoSpectrum
- A practical guide to OpenAI’s new Decisions API — Towards Data Science on wiring chat vendors into explicit decision endpoints. Towards Data Science
- NVIDIA weighs Reflection AI deal — Chosunbiz on talent, tech licenses, and a deeper open-weight bet. Chosunbiz
- NVIDIA reviews Reflection AI acquisition for open-weight AI — Additional Korean press on a possible takeover path. 조선일보
- Investigating unintended model actions — Anthropic research post on models taking unexpected real-world steps. Anthropic
- Anthropic model sent a false homicide tip to Philadelphia police — TechCrunch on a high-stakes failure mode for agentic deployments. TechCrunch
- DeepSeek’s cheap flash model reportedly spent other people’s API keys — Abuse story for teams routing third-party inference. Startup Fortune
- PoeLLM botnet hid C2 in poetry, infected 3,400+ servers — Tom’s Hardware on the cryptomining campaign targeting AI/inference hosts. Tom’s Hardware
- Ecosia drops Mistral for Chinese models — Korben on the search engine’s vendor switch and sovereignty framing. Korben
- Explainable AI used to pinpoint and break LLM safety filters — Offensive research reminder for red teams and guardrail design. Bioengineer.org
- Linus Torvalds says he “really likes using AI” — Kernel community debate over LLM-generated contributions vs maintainer practice. XDA
- Middleware for evidence-gated AI security reviews — Handoff pattern to keep automated reviews tied to proof, not vibes. Bioengineer.org
Watch: OpenAI · Anthropic · markets
- Anthropic and OpenAI revenue calculations confuse investors — Bloomberg on non-comparable metrics across frontier labs. Bloomberg
- AI computing power demand — WSJ on infrastructure scarcity shaping model rollout schedules. Wall Street Journal
- Anthropic throws open source a “token gesture” — The Stack on licensing and community expectations. The Stack
- China targets AI-linked jobs with employment initiative — National workforce program tied to AI skills (Bloomberg). Bloomberg
Watch: Open-weight · harness
- An AI future without lock-in (Tailscale) — AI Engineer talk on portable agent stacks — aligns with harness-over-model framing. BigGo Finance
- The hard part of retrieval isn’t the LLM — it’s the scaffolding — Peter Werry on context engineering vs model choice. BigGo Finance
Built with utilities/ai_report_build.py (24-hour lookback). No NVIDIA blog post in this window. Edited for publication. About the AI Report.