硅谷主要科技企业正在大规模押注人工智能递归自我改进（RSI）技术的发展。[1]谷歌2024年在AI数据中心、芯片和基础设施上的总资本开支达1950亿至2050亿美元，明年还将进一步增加。[1]Google DeepMind首席战略官Jasjeet Sekhon表示，整个科技巨头的投资本质上都在押注RSI，其投资规模已经超过阿波罗登月计划和曼哈顿工程。[1]这被业内人士视为人类文明最大的科学赌局。[1]
在RSI能力演进方面，多位研究人员给出了乐观预期。[1]Chamath Palihapitiya预言接下来18个月RSI将迅速提升AI能力。[1]DeepMind研究副总裁Oriol Vinyals和OpenAI联创Wojciech Zaremba判断真正的RSI"很可能在2027或2028年出现"。[1]Anthropic已展示了相关进展，其九个Claude智能体用时800小时、花费约1.8万美元，追回了97%的性能差距。[1]GPT-5.6 Sol经Codex改写生产环境GPU内核，端到端推理成本下降了20%。[1]
然而这一技术方向也面临重大风险。[1]Sekhon警告称，任何一个人仅通过自然语言与模型对话，就能设计出致命病毒或蛋白。[1]目前RSI仍未完全实现，存在"AI气穴"风险，同时还需应对网络安全和生物安全等极端威胁。[1]
Silicon Valley investors and technology executives are predicting the imminent arrival of an era of recursive self-improvement (RSI) in artificial intelligence, characterizing it as humanity's largest scientific wager.[1] Google, Anthropic, OpenAI, and other major tech companies are channeling hundreds of billions of dollars into AI research to advance the capability of AI systems to autonomously optimize algorithms and design chips.[1]
Google's 2024 capital expenditures on AI data centers, chips, and infrastructure totaled between $195 billion and $205 billion, with further substantial increases expected next year.[1] Jasjeet Sekhon, chief strategy officer at Google DeepMind, asserts that the investment strategies of tech giants collectively represent a bet on RSI, with the scale now exceeding both the Apollo moon landing program and the Manhattan Project.[1] Venture capitalist Chamath Palihapitiya predicts that RSI capabilities will advance rapidly over the coming 18 months.[1] Oriol Vinyals, vice president of research at DeepMind, and Wojciech Zaremba, OpenAI co-founder, estimate that genuine RSI is "likely to emerge in 2027 or 2028."[1]
Despite progress in AI self-optimization capabilities, RSI remains incomplete, and the field faces risks including what experts term an "AI cavitation" phenomenon alongside extreme security and biological safety concerns.[1] Anthropic's nine Claude AI agents demonstrated notable performance recovery, closing 97 percent of a performance gap in approximately 800 hours at a cost of roughly $18,000.[1] GPT-5.6 Sol, after code rewriting through Codex, achieved a 20 percent reduction in end-to-end inference costs on production environment GPU kernels.[1] Sekhon has warned that any individual conversing with an AI model through natural language alone could potentially design a deadly virus or protein.[1]
