• AI KATANA
  • Posts
  • China pitches a global AI cooperation body

China pitches a global AI cooperation body

Also: Hinton calls for shared AI safeguards at WAIC

Hello!

China used WAIC in Shanghai to pitch a new multilateral AI-governance body and flaunt Huawei’s CloudMatrix 384 super-cluster, while Geoffrey Hinton urged shared guardrails to avoid “fragmented futures.” Across the Pacific, Tesla quietly rolled out supervised robotaxi rides around San Francisco and Oakland, stoking the robo-mobility debate. On policy, Washington’s DOGE unit began testing an LLM that flags 100,000 federal rules for possible repeal, even as Microsoft slipped a “Copilot Appearance” makeover into Windows previews. Investors stayed busy: Korean web-3 outfit Story’s Poseidon landed a16z seed capital, and ex-U.S. Navy founders raised cash to pipe submarine data into AI inference engines. Finally, away from the boardrooms, Dubai teased an AI-designed dining experience—proof that silicon chefs may soon plate your lunch. Scroll on for the full slices.

Sliced just for you:

  • 🌐 China proposes global AI cooperation organisation

  • 📡 Huawei debuts CloudMatrix 384 to rival Nvidia

  • 🚕 Tesla starts supervised robotaxi service in Bay Area

  • 🧠 Hinton calls for shared AI safeguards at WAIC

  • 👥 Chat-style AI becomes a lifeline for neurodivergent users

  • 🔬 AI speeds up nuclear-materials analysis for researchers

Premier Li Qiang unveiled a plan for an international organisation headquartered in Shanghai to harmonise AI standards, bridge chip-access gaps for the Global South, and blunt technology “monopolies.” The proposal, tabled at WAIC, includes a voluntary governance framework and a membership drive targeting governments, labs and firms by year-end. Observers see Beijing positioning itself as convener-in-chief, counter-balancing U.S.–led blocs while courting emerging-market allies with infrastructure pledges and model-sharing incentives.

On the same stage, Huawei rolled out CloudMatrix 384—linking 384 Ascend 910C chips in a “super-node” fabric the company says can eclipse Nvidia’s GB200 NVL72 on throughput and latency. Early benchmarks from SemiAnalysis suggest the design’s custom interconnect shaves inter-GPU hop times, albeit with slower single-chip FLOPs. The system is already live on Huawei Cloud and will ship to domestic hyperscalers this quarter, underscoring China’s sprint to self-reliance under U.S. export curbs.

Select FSD-supervised robotaxis began ferrying internal staff and invite-only riders between San Francisco and Oakland. Vehicles operate on geofenced routes with a safety driver monitoring a Level-4-style stack; user feedback is feeding a public launch roadmap slated for Q4. Analysts say the pilot is critical proof-of-concept before August’s planned fully driverless unveil.

AI pioneer Geoffrey Hinton told delegates that siloed national datasets and divergent standards could “trap humanity in incompatible bubbles.” He advocated open-weight research exchanges and joint red-teaming between U.S., EU and Asian labs, framing alignment as “the next CERN.”

A feature report highlights how tools like GPT-based assistants help autistic and ADHD users draft emails, rehearse conversations and filter sensory overload. Interviewees say AI offers judgement-free coaching, yet clinicians caution against over-dependence and privacy gaps.

Researchers demonstrated a model that combines hyperspectral imaging with large-token sequence analysis to map isotope composition in minutes—cutting lab time by 90% and aiding non-proliferation inspections.

🛠️ AI tools updates

Insiders testing the 24H2 Canary build discovered a new panel that lets users pick friendly faces, tones and stylistic quirks for Copilot responses. The feature hints at deeper personalization layers ahead of Microsoft’s fall release push.

An internal slide deck shows the Department of Government Efficiency using a proprietary model to parse 200,000 rules, flagging half for potential repeal to hit President Trump’s “10-for-1” mandate. Civil servants worry about legal accuracy, but early pilots at HUD cut review time from weeks to hours.

💵 Venture Capital updates

Founded by former U.S. submarine commanders, Spear AI is building orchestration layers that turn sonar and acoustic logs into real-time tactical insights. Proceeds will expand GPU clusters and secure SBIR contracts with the Navy’s Undersea Warfare Center.

The Korean startup is creating a decentralized, IP-cleared data marketplace for frontier-model training. Funds will hire 30 engineers, launch an incentive token, and open a Singapore hub to court SEA contributors.

🫡 Meme of the day

⭐️ Generative AI image of the day