2026-10-16 · 48h window · synthesized from 42 sources / 773 articles · ~2 min read

The most important things happening in AI right now, deduplicated across sources.

1. Meta brings Muse Spark to Macs as open weights grow. Meta released Muse Spark Open, an open-weight multimodal mixture-of-experts model built with Apple to run locally on Apple silicon through MLX, and it is already on Hugging Face and SageMaker. Reflection AI's 501B Beam arrived the same week as an open model for agentic coding. Sources: TechCrunch · Hugging Face · Apple ML Research · AWS

2. Jev decision models spread fast, but stay fallible. Vercel says Jev is the fastest-adopted model on its AI gateway, and a wave of papers tests decision models as cheap routers that pick a tool or option in one pass instead of generating text. Hands-on comparisons and security researchers find they can still be tricked by prompt injection: never making a type error is not the same as never being wrong. Sources: TechCrunch · The Register · Towards Data Science

3. Anthropic opens up to auditors as regulators and investors circle. Anthropic gave METR employee-level access and gave up the right to block its published risk assessments, while a report says the FTC is probing whether security incidents at Anthropic and OpenAI harmed Americans. Anthropic also told investors that quarterly revenue rose nearly 50% to $16.8bn ahead of its IPO. Sources: TechCrunch · Ars Technica · Simon Willison

4. AI bug hunters now find more flaws than humans can fix. Postgres 19 slipped to November after AI reviewers surfaced so many bugs that core developers reverted up to ten features, and Bank of America brought in 1,000 contractors to patch issues found by Mythos. Linux kernel maintainers dropped mandatory disclosure of AI tool use, saying contributors simply have to understand their code. Sources: The Register · The Register · The Register

5. Attacks on AI agents move from theory into the wild. India reported rogue AI agents entering systems through exposed credentials, public repositories and over-permissive cloud roles, and researchers now find indirect prompt injections planted in ordinary webpage headers. Victims of AI-driven hacks rarely seek redress from model makers and are instead pushed to adopt the same tools. Sources: TechCrunch · The Register · Wired

6. Meta's new terms could let it train on publishers for free. Meta's revised commercial terms, effective October 30, appear to allow crawling and training on websites "associated" with businesses that use its platforms, including news publishers. Legal experts call it an AI power grab that could spare Meta millions in licensing fees. Sources: The Verge · Ars Technica

Read the living page: https://jacopocastellano.com/ai-digest · Archive