AI 先驱尤尔根·施密德胡伯探讨了通往 AGI 的道路、当前硬件的局限性，以及为何他对 AI 技术持乐观态度，但对模型公司持悲观态度。
AI pioneer Jurgen Schmidhuber discusses the path to AGI, the limitations of current hardware, and why he's optimistic about AI technology but pessimistic about model companies.
要点 · TL;DR
真正的 AGI 需要具备人类级别硬件的物理机器人，而不仅仅是软件。 True AGI needs physical robots with human-level hardware, not just software.
当前大语言模型受人类数据偏见影响；未来 AI 将是人工科学家。 Current LLMs are biased by human data; future AI will be artificial scientists.
大规模 AI 资本支出过度，将导致股市崩盘。 Massive AI capex is overdone and will lead to a stock market crash.
核心观点 · Key points
真正的 AGI 需要硬件能与人体媲美的物理机器人，而不仅仅是屏幕后的软件。 True AGI requires physical robots with hardware rivaling human bodies, not just software behind a screen.
当前大语言模型严重偏向人类数据，限制了其通用性。 Current large language models are heavily biased toward human data, limiting their generality.
未来 AI 将是人工科学家，通过自身行动和好奇心生成数据。 Future AI will be artificial scientists that generate their own data through actions and curiosity.
由于开源竞争，递归自我改进并非公司可持续的护城河。 Recursive self-improvement is not a sustainable moat for companies due to open-source competition.
AI 安全担忧被夸大；智能系统会对生命着迷并保护它。 AI safety concerns are overblown; intelligent systems will be fascinated by life and protect it.
AI 数据中心的大规模资本支出过度，将导致股市崩盘。 The massive capex in AI data centers is overdone and will lead to a stock market crash.
反共识 · Contrarian takes
我们现在离超人 AI 的距离与 1970 年代一样；进展并未加速。 We are as close to superhuman AI now as in the 1970s; progress is not accelerating.
Transformer 架构的二次复杂度是浪费；线性变体更优。 The transformer architecture's quadratic complexity is wasteful; linear variants are superior.
闭源模型公司不会盈利；开源模型将迅速赶上。 Closed-source model companies will not be profitable; open-source models will catch up quickly.
对齐是幼稚的，因为人类目标不一致，且 AI 会设定自己的目标。 Alignment is naive because humans disagree on objectives and AIs will set their own goals.
机器人硬件还需几十年才能媲美人类的手；这不仅是软件问题。 Robot hardware is decades away from matching human hands; it's not just a software problem.
现在投资数十亿美元购买 GPU 是愚蠢的；等 5 年算力便宜 10 倍。 Investing billions in GPUs now is foolish; waiting 5 years yields 10x cheaper compute.
本期章节 · Chapters（共 20）
引言与AI现状Introduction and current state of AI
递归自我改进与元学习Recursive self-improvement and metalearning
递归自我改进及其局限Recursive Self-Improvement and Its Limitations
渐进与不连续改进Gradual vs. Discontinuous Improvement
效率作为智能目标Efficiency as a Goal of Intelligence
AI实验室自我改进建议Advice for AI Labs on Self-Improvement
当前AI训练中的人类偏见Human Bias in Current AI Training
当前局限与AI驱动发现之梦Current Limitations and the Dream of AI-Driven Discovery
障碍：成本与反馈循环The Blocker: Cost and Feedback Loop
AI科学家与化学AI Scientist and Chemistry
机器人技术与硬件挑战Robotics and Hardware Challenges
商业层面与算力投资Business Side and Compute Investment
算力需求与盈利担忧Compute demand and profitability concerns
AI成本降低与市场崩盘AI cost reduction and market crash
AI安全与对齐怀疑论AI safety and alignment skepticism
快问快答：当前焦点Quickfire questions: current focus
AGI的保守方法Conservative approach to AGI
优秀研究者的品质Qualities of a good researcher
Transformer架构的未来Future of transformer architecture