🤖 AI Daily · Jul 11, 2026

Apple sues OpenAI for trade secret theft; Baidu AI Day unveils 4 Dazi updates; Claude Code desktop gets in-app browser; Zuckerberg denies Meta compute surplus: renting is more profitable; Ant Group open-sources SGLang inference framework; Claude Code v2.1.206; Unitree G1 humanoid performs first live minimally invasive surgery; Xiaohongshu unveils PIPO architecture; DeepSeek-V4 Flash lands on AMD MI355X GPUs; How Cognition trusts Claude Fable 5 to work overnight.

1. 1. Apple Sues OpenAI for Trade Secret Theft

   Apple files lawsuit alleging OpenAI stole trade secrets and breached contracts. Claims OpenAI execs directed former Apple employees to steal confidential info during hiring.

2. 2. Baidu Dazi Unveils 4 Major Updates

   Baidu Dazi updates: personal version with browser调用 & smart routing (-20% latency), creator suite, enterprise edition, and Dazi Alliance.

3. 3. Claude Code Desktop Gets In-App Browser

   Claude Code desktop now has an in-app browser. Claude can view documents, designs, and websites, interacting like a local dev server. Sandboxed and configurable.

4. 4. Zuckerberg Denies Meta Compute Surplus

   Zuckerberg denies Meta has excess compute, says internal demand is at full capacity. But acknowledges renting out AI infra is more profitable.

5. 5. Ant Group Open-Sources SGLang Inference Framework

   Ant Group open-sources SGLang, a high-performance inference serving framework for LLMs and multimodal models on GitHub.

6. 6. Claude Code v2.1.206 Released

   Claude Code v2.1.206 adds directory suggestions for /cd, /doctor to trim CLAUDE.md, and improved /check.

7. 7. Unitree G1 Humanoid Performs First Live Minimally Invasive Surgery

   UCSD team uses Unitree G1 humanoid robot to perform laparoscopic cholecystectomy on live pigs - first standard minimally invasive surgery by a humanoid robot.

8. 8. Xiaohongshu Unveils PIPO Architecture

   Xiaohongshu proposes PIPO architecture: compresses 2 tokens into 1 latent on input, expands on output, halving input length and doubling per-step output.

9. 9. DeepSeek-V4 Flash Lands on AMD MI355X GPUs

   DeepSeek-V4 Flash RL training now supported on AMD Instinct MI355X GPUs via Miles framework on ROCm. The 284B parameter MoE model runs on AMD hardware.

10. 10. How Cognition Trusts Claude Fable 5 to Work Overnight

    Cognition's SVP says Devin tested nearly all Claude models; Claude Fable 5 is the first trustworthy enough to run unsupervised through the night.