Jürgen Schmidhuber 对 AI 现状的洞见Jürgen Schmidhuber on the State of AI
Jürgen Schmidhuber,被誉为 AI 之父,他认为从宇宙视角看,我们距离 AGI 并不遥远,但在物理世界中,机器人硬件远逊于人体,这是真正 AGI 的主要障碍。他对当前大规模数据中心投资持怀疑态度,认为计算成本会快速下降,导致许多公司面临巨大损失,并预测可能出现股市调整。尽管对技术乐观,但他对依赖预训练模型的公司盈利能力持悲观态度。
在递归自我改进(RSI)方面,他强调当前方法是早期工作的扩展版本,通过元学习实现效率提升。Schmidhuber 还提出人工科学家概念,通过好奇心驱动的自实验收集数据,减少对人类数据的依赖。他对 AI 安全担忧较轻,认为赋予 AI 自主设定目标的能力更重要,且它们会像科学家一样保护生命来源。
Schmidhuber 认为未来在于物理 AI 和自我复制机器人社会,这将开启太阳系殖民等宏大愿景。
在递归自我改进(RSI)方面,他强调当前方法是早期工作的扩展版本,通过元学习实现效率提升。Schmidhuber 还提出人工科学家概念,通过好奇心驱动的自实验收集数据,减少对人类数据的依赖。他对 AI 安全担忧较轻,认为赋予 AI 自主设定目标的能力更重要,且它们会像科学家一样保护生命来源。
Schmidhuber 认为未来在于物理 AI 和自我复制机器人社会,这将开启太阳系殖民等宏大愿景。
Jürgen Schmidhuber, widely regarded as a father of AI, believes that from a cosmic perspective we are not far from AGI, but in the physical world, robot hardware lags far behind the human body, posing the main barrier to true AGI. He is skeptical of the current massive data center investments, arguing compute costs will drop rapidly, leading to significant losses for many companies and potentially a market correction. While optimistic on the technology, he is pessimistic about the profitability of companies relying on pretrained models.
On recursive self-improvement (RSI), he notes current approaches are scaled versions of early work using meta-learning for efficiency gains. Schmidhuber advocates for artificial scientists driven by curiosity to generate their own experimental data, reducing reliance on human web data. He is notably less worried about AI safety, emphasizing the importance of letting AIs set their own goals, as they will act like scientists motivated to protect sources of interesting patterns and life.
He sees the future in physical AI and self-replicating robot societies that could colonize the solar system.
查看原文 →
On recursive self-improvement (RSI), he notes current approaches are scaled versions of early work using meta-learning for efficiency gains. Schmidhuber advocates for artificial scientists driven by curiosity to generate their own experimental data, reducing reliance on human web data. He is notably less worried about AI safety, emphasizing the importance of letting AIs set their own goals, as they will act like scientists motivated to protect sources of interesting patterns and life.
He sees the future in physical AI and self-replicating robot societies that could colonize the solar system.