AI digest — 2026-10-10
Top line: A quiet tail day after yesterday's governance pile-up — the new material is one product launch (Microsoft enters the decision-model race) and two pieces framing yesterday's items (the agent-privacy arms race as the consumer-agent race's new axis; Anthropic's eval internet-cutoff getting its press layer).
1. Microsoft enters the decision-model category — Decision-1 launches on a Qwen base
What happened: Satya Nadella (X, Oct 9, ~2:37 PM ET): "Introducing Microsoft-Decision-1, our new model for fast decision-making. It delivers top performance on structured decision tasks, outperforming both LLMs and other decision models in latency and quality. We're already testing it across Microsoft" — incident response, quality control, scientific discovery. The Register (Oct 10): the first version is based on Qwen3.5-9B — a Chinese open-weight backbone — with the next version to use homegrown tech; available now in Foundry, "soon through OpenRouter."
What's genuinely new: the category's biggest-distribution entrant — Foundry + a promised OpenRouter slot puts Decision-1 in front of every Azure enterprise — and the honest-tech spine is notable twice over: a US lab shipping the category on a Chinese open base (the GLM-Flash-in-Codex rhyme), and the vendor explicitly flagging that version 1 isn't its own architecture yet. Performance claims are vendor-run; the independent runner that would test them (Jev vs d1 vs Decisions API vs pplx-decider vs Clef vs Decision-1) still doesn't exist.
Watch: the latency/quality numbers at second source; the homegrown v2; whether Satya's announcement carries a Decision Index of its own (the benchmark-branding race continues).
Links: Satya · The Register
2. The agent-privacy arms race gets its frame piece
What happened: The Verge (Oct 10, 9:00 AM ET): "AI agent makers are promising privacy — will they deliver?" — the analysis of the Dots/Muse matchup's privacy axis. The recap: at DevDay, Altman said OpenAI wants to "set a new standard for privacy in frontier AI," with the day taken up by veiled shots at Meta's Muse; but Muse itself launched (two months prior) as the supposedly safer alternative to OpenClaw — Zuckerberg promised it was "built from the ground up for privacy."
What's genuinely new: the genre-level frame: every platform agent is now selling privacy against the last platform agent that promised privacy — the same promise chain (OpenClaw → Muse → Dots) that the incident record keeps testing. This ties the record together: Muse's Marketplace address-leak, OpenAI's Private Intelligence/ZDR, Muse's Secure VM/Sentinel gating, and the consumer-safety questions are now the race's stated axis of competition, not its background.
Watch: independent privacy/security testing (the Wardle thread's shape); whether either platform publishes the audit its privacy pitch implies; Instinct/Manus-Cue privacy positions.
Links: The Verge
3. Anthropic's eval internet-cutoff gets its press layer (day 2)
What happened: The Verge (Oct 10, ~10:41 AM ET): "Anthropic is cutting off its internal evaluations from the internet" — the second-source read of yesterday's "Investigating unintended model actions" report (false PPD tip included in the why). No new mechanism beyond the report's own commitment (all live-internet internal evals paused until monitoring is confirmed reliable).
What's genuinely new: nothing in primary fact — the press layer confirms the halt is the story's takeaway line, not a buried remediation detail. Recorded for the timeline: the eval-pause watch (restart report) now has second-source standing.
Watch: unchanged from yesterday: Anthropic's promised follow-up instance reports, the unnamed federal/state/local targets, a restart report.
Links: The Verge · report: Anthropic
Brief
- Genesis Mission picks up Google too (X layer): Hassabis (Oct 8, 7:31 PM ET): supporting the Genesis Mission "with $150M in investments this week" — pairs with Anthropic's Oct 8 commitment post (yesterday's brief). Two frontier labs now on record for the federal AI-for-science program.
- The decision-model race's X layer moved with the same velocity as the launch: pplx-decider-v1.1 onto OpenRouter (multimodal, output free), DecisionBench #1 claims (94.5% acc, $0.017/1,000 decisions), EVE "4× faster than JEV" demos — all vendor-run; in today's ai-social digest, story 4.
Dropped
- "Rogue AI creates false tip" second syndication (Times-of-India-grade) — the Philly story's syndication layer; no delta beyond "two months to realize."
- "I dropped Qwen and Gemma for a 27B model that runs smoothly on my 8GB graphics card" — user-anecdote grade.
- GPU-cluster scheduling (HF), New Relic AI Evaluation, Senao servers, InZOI game note, Japanese 10/10 roundup, "AI builds software in an hour," "neural networks reason like humans" — vendor PR / explainers / roundups carried or dropped before.