Meta pushed consumer agents with Muse while Anthropic published self-replication and distillation-defense research. Infra money and risk stayed loud: data-center build costs, capital-crunch warnings, and Microsoft expanding Azure Maia. On-device and open stacks: LiquidAI LFM2.5 on Hugging Face, plus MCP and LiteLLM security findings.
Today
- Welcome to October 13, 2026 — AWG’s daily Innermost Loop firehose. The Innermost Loop
- Meta Muse Wonder — Meta research post on the Muse agent stack. Meta AI Research
- Wired: Muse AI agent boundaries — Product behavior stress-test; Meta’s response framing. Wired
- Anthropic: self-replicating AI — Research note on replication dynamics and evals. Anthropic
- Anthropic: distillation defenses — Hardening against weight/behavior extraction. Anthropic
- GLM Legal (z.ai) — Open-leaning stack courting U.S. legal workflows. z.ai
- Fight over what independent AI reviews make public — Industry calls for review vs disclosure limits. Reuters
- Chinese mathematicians unite behind AI research use — SCMP on research-community alignment with AI tools. South China Morning Post
- Unitree founder: GPT-6 may unlock robotics next year — OpenAI roadmap talk tied to embodied AI. South China Morning Post
- Bloomberg: AI data-center build costs — Feature on what gigawatt-scale buildouts actually cost. Bloomberg
- How the AI boom is fueling global financial risks — Macro spillover framing for the capex wave. Bloomberg
- Intel and Broadcom CEOs: capital crunch could bottleneck AI — Funding, not just silicon, as the gate. CNBC
- NTT to build first Anthropic data center in Japan — Regional capacity for Claude stack. Nikkei Asia
- Microsoft expands Azure Maia to multi-gigawatt deployments — Custom silicon path at cloud scale. Microsoft
- LiquidAI LFM2.5-24B-A2B: fast private on-device agents — Hugging Face write-up for the compact MoE-style release. Hugging Face Blog
- AI-assisted Linux patches cut kernel build times — Path toward ~10-second compiles without a ramdisk. Tom's Hardware
- Epic Games delays Disney project, asks for more AI — Studio schedule pressure meets generation tooling. Bloomberg
- IMF: AI could hit ~40% of jobs, young workers first — Labor-market warning making the rounds. Decrypt
- China’s AI energy future — Bloomberg graphics on power as the binding constraint. Bloomberg
- AI assistants in nuclear power plants — IEEE Spectrum on high-stakes copilots. IEEE Spectrum
- WSJ: AI has a god problem — Religion and meaning narratives from Vatican to Baptist pulpits. Wall Street Journal
From NVIDIA
- UK ecosystem: from enterprise to Earthshot with NVIDIA — Deployment story across UK AI buildout. NVIDIA Blog
- European cities deploy NVIDIA-powered AI agents — Urban operations use cases. NVIDIA Blog
- PC app ecosystem ready for NVIDIA RTX Spark — Local AI software stack for Spark laptops. NVIDIA Blog
Watch: Claude / Anthropic
- Self-replicating AI + distillation defenses — Pair of Anthropic research posts (also in Today). Anthropic
- NTT Japan data center for Anthropic — Capacity move in Asia. Nikkei Asia
Watch: OpenAI
- GPT-6 and robotics (Unitree founder) — Embodied timeline speculation tied to next frontier model. South China Morning Post
Watch: Elon / xAI
- Elon Musk on X — Flagged in today’s AWG edition. X
Watch: Open-weight · security · infra
- 253 vulns across 343 MCP server / agent tools — Agent-tool surface area under audit. GBHackers News
- LiteLLM flaws: gateway hijack and provider-key theft — Low-priv path to high-value secrets. IT Security Guru
- Modular MAX tops vLLM by 5.5× on Kimi K2.5 — Inference-engine horse race continues. shattered.io
- Mistral’s China strategy as a European bet — fDi on competitive positioning. fDi Intelligence
- Open models: benefits vs risks — MIT Technology Review framing piece. MIT Technology Review
- AMD Ryzen AI Halo 192GB vs DGX Spark speed — Local large-model RAM vs NVIDIA Spark throughput. Tech Times
- Microsoft Decision-1 joins the decision-model cloud race — Scorer/router models shipping on major clouds. Tech Times
Built with utilities/ai_report_build.py (24-hour lookback). Edited for publication. About the AI Report.