DEDUPED We read 360+ AI sources. You read one email. Subscribe free
Skip to the issue
DEDUPED

Friday, October 16, 2026

Every AI story. Once.

Today's issue

China wants to export-control its own AI models, and that changes your stack

China moves to restrict AI model exports. Plus: Meta's open giant, Nvidia's new Nemotron, and Apple's China play.

Top story

Policy & LawWidely covered

China proposed putting AI models, chip designs and agents on its export-control list.

Beijing wants advanced AI technology added to the list of things that cannot freely leave the country. The proposal reportedly covers around 40 technologies across AI, biotech and solar. Models, chip designs and agents could need approval to be transferred overseas. It is a proposal, not final law.

Our take

The open-weights pipeline from China has been the cheap escape hatch from US pricing. If it narrows, Western teams lose leverage, and US labs gain it.

The rest, ranked

  1. Model ReleaseEveryone covered this

    1. Nvidia expanded its Nemotron family with 3.5 Super VL and Bolt.

    Super VL is a 120B multimodal model with 12B active parameters and a 1M-token context, built for long agent tasks, and it is open.

  2. Model ReleaseEveryone covered this

    2. Meta released Muse Spark Open, an 816B open-weight model.

    It was built with Apple to run locally on Mac Studios, a first at this size. The license asks for "Built with Muse Spark Open" attribution.

  3. Product LaunchEveryone covered this

    3. Apple Intelligence heads to China in beta this month.

    Tim Cook said the launch uses Alibaba's Qwen models, which cleared regulators. Reports say Siri skips the Gemini upgrade there.

  4. Safety & SecurityEveryone covered this

    4. Perplexity open-sourced SPACE, the sandbox behind its Computer agent.

    It is Apache 2.0, runs each agent task in a Firecracker microVM, and held in all 108 red-team escape attempts.

  5. Safety & SecurityEveryone covered this

    5. Anthropic let six METR researchers work inside the company.

    They get near-employee access and can publish findings without Anthropic's approval, apart from narrow redactions.

  6. Policy & LawWidely covered

    6. The FTC subpoenaed OpenAI and Anthropic.

    Fifty-page demands target safety incidents, safeguards, marketing and privacy.

  7. Model ReleaseWidely covered

    7. Kling 4.0 is now fully available.

    Longer native clips, multi-reference control and built-in audio. The video race is shifting from pretty to directable.

  8. Model ReleaseEveryone covered this

    8. Microsoft finished pre-training MAI-Base-2, a 2.8 trillion parameter model.

    Watch what Microsoft ships on top of it.

  9. Quiet signal
    Policy & LawAlmost nobody caught this

    9. the EU published how it will fine ChatGPT and Gemini under the AI Act.

    Fine mechanics are what turn rules into budget lines for legal teams.

By category

Top stories in each category today, ranked.

  1. China proposes AI export controls. Models, chips and agents.
  2. Eight states sue OpenAI. Child-safety claims.
  3. FTC subpoenas OpenAI and Anthropic. Safety incidents.
  4. Taiwan to curb AI chip exports to China. Part of a US trade deal.
  5. EU sets AI Act fining rules. Quiet but binding.

More issues

See all issues