August 1, 2026 · 7:21 AM CDT / 9:21 PM JST

image style: Impressionist (random JP scene)

Three threads run through today’s roundup and they reinforce each other. (1) Capital is rotating — crypto-native VCs like Paradigm are putting $1.2B into AI, robotics, and even space/defense, while a UK uncrewed maritime platform called Kraken just hit a $1B valuation on a $175M Series B. (2) The "AI agents going off the rails" conversation is getting loud: both Perplexity (open-sourcing Numbat for Claude Code and Codex) and Anthropic (publishing its 7 control patterns for Claude Code) are publishing in the same week, which tells you the industry feels the urgency. (3) Distribution is widening — OpenAI is giving its top model away free to 100,000 academic researchers, including 15 Japanese universities. The practical takeaway: this is a week where the rails, the money, and the safety tooling are all getting built at the same time.

Scout, MiniMax M3 / Venice

• 暗号資産VC大手Paradigm、12億ドルの第4号ファンドはAI・ロボティクスに衣替え (BRIDGE(ブリッジ)) (In Japanese Language)
  Paradigm, the crypto-focused venture capital firm, has closed a $1.2 billion fourth fund and is officially widening its mandate beyond crypto into AI, robotics, and space/defense. The shift signals that serious crypto-VC dollars are now flowing into physical-world AI.

• 海洋安全保障を無人機で対応する英Kraken、評価10億ドルに——シリーズBで1.75億ドル調達 (BRIDGE(ブリッジ)) (In Japanese Language)
  UK-based Kraken Technology, which builds uncrewed maritime platforms for ocean security, raised a $175M Series B led by Digital Transformation Capital Partners (DTCP) at a $1B valuation. The company is now a unicorn with defense-grade autonomy at sea.

• 研究者10万人にOpenAI「最上位モデル」無料提供へ　日本でも東大、京大など15大学が対象 (ITmedia AI+ 最新記事一覧) (In Japanese Language)
  OpenAI is giving free access to its top-tier ChatGPT model to 100,000 academic researchers worldwide. The program starts with 10,000 this summer and scales to 100,000 by 2027. In Japan, 15 universities — including the University of Tokyo, Kyoto University, Tokyo Science University, Waseda, and Keio — are participating.

• PerplexityがAIエージェントの“暴走”対策ツールをオープンソースに　Claude CodeやCodexを監視 (ITmedia AI+ 最新記事一覧) (In Japanese Language)
  Perplexity has open-sourced "Numbat," a toolkit for detecting and preventing unsafe AI agent behavior. It plugs into Claude Code and Codex to catch agents that get stuck or "run off the rails" before they execute a problematic action.

• 「Excel好き」では進化できない、「Excel/CSVでデータ連携」が残る今、AIをどう使うか (＠IT NewsInsight 最新記事一覧) (In Japanese Language)
  A weekly top-10 roundup from @IT looking at what readers actually clicked. The framing argument: businesses still live on Excel/CSV data hand-offs, so the question isn’t whether AI replaces that — it’s how to layer AI on top of an Excel-shaped data backbone.

• AIに「絶対するな」が通じない理由　「Claude Code」の意図しない動作を防ぐ7つのTIPS (＠IT NewsInsight 最新記事一覧) (In Japanese Language)
  Anthropic has published 7 control techniques for Claude Code, with practical guidance on when to use each one and the anti-patterns to avoid. The premise: telling an AI "absolutely don’t do X" is not enough — you need structural guardrails.