AI 正着手打造更强的下一代 AI

TL;DR - AI systems are increasingly helping design, code, and evaluate their successors, but today’s tools remain far from fully autonomous recursive self-improvement. Human-defined goals, evaluation criteria, oversight, and physical infrastructure remain crucial constraints.

• GPT-5.3-Codex, Claude Code, and AlphaEvolve illustrate AI-assisted model development, algorithm discovery, chip design, deployment, and evaluation.
• Darwin Gödel Machines can modify their coding-agent software, while AI Scientist targets automated ideation, experimentation, paper writing, and review.
• Full recursive self-improvement requires autonomous idea generation, evaluation, and process refinement—not merely better outputs—and current systems do not meet that standard.
• Researchers expect friction from complex architectures, high costs, tacit knowledge, and real-world operations, making sustained human–AI collaboration more plausible than an imminent uncontrolled intelligence explosion.