AI Briefing, March 31 · Issue #163
Editorial standards and source policy: Editorial standards, Team. Content links to primary sources; see Methodology.
## 🔍 Key Insights
**Embodied AI** is rapidly moving from simulation to real-world deployment. Hardware players like **AAC Technologies** and **Seeed Studio** are deeply integrating into robotics’ perception and actuation layers. Meanwhile, **Ollama** is aggressively enhancing local inference capabilities—adding native support for **MLX**, **NVFP4**, and advanced caching optimizations—solidifying Apple Silicon as a new powerhouse platform for AI development [1][2][9][15][13]. At the same time, **supply chain security risks** are escalating: the severe attack on **axios**, coupled with the rise of “vibecoding,” has triggered broad industry reflection on the fragility of modern development practices [4][18].
## 🚀 Major Updates
- **First Global Embodied AI Developer Conference & “Juliang Plan” Hackathon features live robot vs. robot PK** [1]: Embodied AI models were deployed end-to-end—and stress-tested in adversarial scenarios—for the first time in real physical environments.
- **AAC Technologies unveils a humanoid robot acoustic perception solution** [2]: Leveraging its deep expertise in consumer electronics acoustics, AAC delivers a tightly integrated hardware-software system for multimodal environmental sensing.
- **Ant Group’s Lingbo open-sources the 2.71TB spatial intelligence dataset *LingBot-Depth-Dataset*** [3]: Contains 3 million RGB-D image pairs, specifically engineered to tackle persistent challenges in transparent and reflective object recognition.
- **Ollama natively integrates Apple’s MLX framework; Qwen3.5-35B achieves dramatically reduced first-token latency on M5 chips** [9][16][17]: Local LLM inference speed and energy efficiency leap forward by a generation.
- **Ollama adds NVFP4 format support and refines agent caching mechanisms** [15][13]: Improves consistency across NVIDIA GPU production environments and cuts memory overhead in agent-driven workflows.
- **A Meta intern (of Chinese origin) proposes HYPERAGENTS—a framework enabling code-level self-evolution of AI agents** [10]: Combines Gödel machines with Darwinian open algorithms, enabling meta-learning–driven autonomous iteration.
- **Popular npm package ‘axios’ suffers a major supply chain attack; Andrej Karpathy urgently warns about unlocked dependency risks** [4]: The “vibecoding” culture is cited as a contributing factor that amplifies systemic vulnerability exposure [18].
- **Bose Quantum secures ¥1 billion Series B funding to mass-produce 1,000-qubit specialized quantum computers** [5]: Accelerating photonic quantum chip R&D, anchored to its “quantum + supercomputing + AI convergence” technical roadmap.
## 🔗 Sources
[1] Robots go head-to-head—in person! This hackathon just leveled up — https://www.bestblogs.dev/article/7c70193c
[2] AAC Technologies reveals its humanoid robot perception solution—signaling accelerated commercial rollout — https://www.bestblogs.dev/article/36080140
[3] 2.7TB of high-quality data trains SOTA spatial intelligence—fully open-sourced, stack-wide — https://www.bestblogs.dev/article/12ea34c2
[4] npm package ‘axios’ hit by major supply chain attack — https://www.bestblogs.dev/status/2038849654423798197
[5] Bose Quantum closes ¥1 billion Series B to industrialize specialized and general-purpose quantum computers — https://www.bestblogs.dev/article/279aa7a4
[9] Ollama adds native MLX support—boosting local LLM inference performance — https://www.bestblogs.dev/status/2038837864340000802
[10] Meta intern (Chinese origin) introduces HYPERAGENTS: