Daily Archive
AI Daily Brief 2026-06-28: OpenAI Hires Apple Hardware Chief, Anthropic Model Limited, Google Caps Meta’s Gemini Access, DeepSeek Launches Accelerator and New Models
Today's AI landscape saw major moves: Apple's Vision Pro chief Paul Meade jumps to OpenAI hardware; Anthropic's Mythos model gets limited US approval for just 100 firms; Google restricts Meta's Gemini access due to compute shortage; DeepSeek unveils DSpark acceleration system and Eagle3 model series; NVIDIA announces Vera Rubin shipping this fall with 10x lower token cost.
Models & Agents
- DeepSeek V4 gets 85% faster inference: Lianhe Zaobao reports significant speedup.
- DeepSeek launches DSpark acceleration system: Aims to reduce inference cost and ease compute bottlenecks (SCMP, Lapaas Voice).
- DeepSeek releases Eagle3 series: Three models based on Gemma4 and Qwen3 with TTT7 test-time training.
- Google makes Computer Use a built-in tool in Gemini 3.5 Flash: Enables native desktop agent capabilities.
- Liquid AI ships LFM2.5-230M: Supports llama.cpp, MLX, vLLM, SGLang, ONNX for on-device inference.
- Baidu open-sources new OCR model: Can process an entire book at once; author suspected to be ex-DeepSeek researcher.
Business & People
- Apple hardware VP joins OpenAI: Paul Meade, head of Vision Products Group, moves to OpenAI's hardware unit.
- Anthropic accuses Alibaba of model theft: Claims 25,000 fake accounts used to extract Claude capabilities.
- Jensen Huang invests in ex-Anthropic founders' startup: ¥1B valuation pre-product, highlighting AI talent premium.
- Broadcom and OpenAI launch Jalapeño AI chip: Market sees AVGO upside.
- Amazon drops Guadagnino's OpenAI film 'Artificial': Director 'not surprised' by cancellation.
Policy & Regulation
- Anthropic Mythos 5 gets limited approval: US Commerce Dept allows 100 US cybersecurity firms access after earlier national security shutdown.
- White House urges OpenAI to restrict GPT-5.6 to vetted partners: Multiple sources confirm.
- GPT-5.6 rate-limited after launch: Overseas supply cuts may benefit Chinese domestic AI firms.
- US limits frontier AI models to Trump-approved customers: 1News reports.
Infrastructure & Partnerships
- Google restricts Meta's Gemini access: FT, WEEX, Phoenix New Media report compute shortage as cause.
- NVIDIA Vera Rubin ships this fall: 10x token cost reduction, HBM4 triple bandwidth, 8 cloud partners.
- Supermicro launches Intel edge AI systems: Up to 367 TOPS with Core Ultra Series 3 and Arc Pro B.
- China Mobile and Huawei deepen 5G-A × AI integration: Aiming for intelligent telecom services.
- Austria lobbies EU to host Anthropic: Sees US regulatory tightening as opportunity.
- Okta first to bring AI agent governance inside FedRAMP: Boosts public sector AI compliance.
- Darktrace joins OpenAI Daybreak cyber program: Transforms security alerts into business impact analysis.
- Alibaba Cloud named leader in Omdia Agentic AI report: Asia & Oceania region.
Other
- AI predicts World Cup teams: Tencent Hunyuan got 29/32 correct, leading MiniMax and iFlytek Spark.
- Anthropic report on global worker behavior: Sleep advice at 5 AM, recipes at 6 PM.
- Chongqing AI Bay Area hosts innovation conference: 'Tongjing' platform debuts.
Why it matters
Today's events highlight three converging themes: intense talent and technology competition (Apple→OpenAI, Huang's investment, Anthropic vs Alibaba); deepening government regulation (White House intervention on GPT-5.6, limited Anthropic approval, US export-style controls); and escalating infrastructure rivalry (Google vs Meta on compute, NVIDIA's Vera Rubin, DeepSeek's accelerator push). These forces are collectively reshaping the global AI industry's trajectory.