AIニュース Nov 2, 2026

今日の注目点

• RSI and weight editingAnthropic's warning that it is close to recursive self-improvement lands alongside research reportedly showing LLMs like GPT-6 Astra can edit other models' weights, suggesting capability gains and direct model-to-model modification are advancing on parallel tracks with unsettled safety implications.
• China AI buildoutChina's proposed export controls on 23 AI technologies and CCID's report that daily token calls exceeded 175 trillion by June point to a buildout shifting from training to inference, just as Meituan's LongCat-2.5 targets W4A4 inference on domestic compute.
• Enterprise AI adoptionGoldman Sachs is restricting Claude Code and GitHub Copilot to engineers after earlier praising 'vibe coding' by non-technical staff, a reversal that contrasts with Apple's planned China beta of Apple Intelligence integrating Alibaba's Qwen into Siri and Writing Tools.
• Benchmarks and data rightsSemiAnalysis's GDPx benchmark, where Opus 5.5 reportedly topped the leaderboard at 12%, arrives as the NYT and other publishers sue Brave's maker over alleged scraping, highlighting how capability measurement and training-data sourcing remain contested.
• AI funding and M&ARobot and data-infrastructure funding continues at scale, with Tutor raising a $75M Series B led by a16z and Beijing's GenRobot AI completing a $100M Series A+, while Saudi Aramco's $2.6 billion deal for 49% of Humain shows state-linked capital entering AI.

AI 生成分析