Gary Marcus:The Road to AI We Can Trust(RSS)
36AI 编辑部评分,满分 100

生成式AI的幻觉、超大规模投入的狂热与世界模型及神经符号AI的路径

2026-05-17 16:02· 91天前· Gary Marcus
AI 导读

近期访谈指出,当前大语言模型存在“幻觉”问题,答案常不可靠。业界对“超大规模”的巨额投入可能陷入非理性狂热。作为替代路径,“世界模型”旨在让AI理解物理规律,“神经符号AI”则尝试结合深度学习与符号推理,以提升可靠性、可解释性与逻辑能力,为下一代AI奠定基础。

Three excellent new interviews

  1. Fabulous deep dive with Brian Greene / The World Science Festival:
  1. Excellent brief but focused interview on the risks of massive bets on hyperscaling, with Zachary Karabell; keynote earlier this week at Web Summit:


2. More technical; also excellent: why we so badly need neurosymbolic AI and world models, and why software verification is more important than ever in the LLM era. A spicy fireside chat hosted by Will Wilson, CEO of Antithesis, from the lively conference Bug Bash 2026, “A conference on extracting reliable software from the slop factory”:

Bonus: Check out YouTuber Husk on using a GenAI-fueled to bargain on his behalf. LLM-powered AI agents are gonna be great! Watch to the end!

Discussion about this post

Author

is there a good set of links on the assumptions hyperscalers have med about depreciation?

May 17Edited

Great stuff, Gary -- you've been busy! (Also, great meeting you at TAIS in Oxford!)

I cover many of these issues (the fundamental cognitive inadequacy of LLM-based chatbots (and why this will never be fixed by scaling), the construction of fully-interpretable world models via continuous learning, and neurosymbolic architecture for AGI) in my paper "Gold-Standard AGI: Outer AGI Superalignment", which I have just submitted to the Springer journal AI & Ethics.

At 214 pages (including 534 references), I admit that my paper may be a little daunting to read -- considering the very high stakes (the future of all humanity for all eternity), I felt that I needed to address every sub-problem as clearly and thoroughly as possible. However, I do believe it’s possible to get at least the gist from reading the Introduction and Conclusion sections alone.

A preprint of the paper (which took me ~8 years to develop) is available here: https://doi.org/10.5281/zenodo.16876832. The next paper in the sequence (on inner AGI superalignment) is already in my head — I just hope it doesn’t take another 8 years! :-)

Ready for more?

来源:Gary Marcus:The Road to AI We Can Trust(RSS) · garymarcus.substack.com

生成式AI的幻觉、超大规模投入的狂热与世界模型及神经符号AI的路径

Gary Marcus:The Road to AI We Can Trust(RSS)·2026-05-17 16:02·91天前·Gary Marcus
AI 导读

近期访谈指出,当前大语言模型存在“幻觉”问题,答案常不可靠。业界对“超大规模”的巨额投入可能陷入非理性狂热。作为替代路径,“世界模型”旨在让AI理解物理规律,“神经符号AI”则尝试结合深度学习与符号推理,以提升可靠性、可解释性与逻辑能力,为下一代AI奠定基础。

原文 · 保持原样,未翻译

Three excellent new interviews

  1. Fabulous deep dive with Brian Greene / The World Science Festival:
  1. Excellent brief but focused interview on the risks of massive bets on hyperscaling, with Zachary Karabell; keynote earlier this week at Web Summit:


2. More technical; also excellent: why we so badly need neurosymbolic AI and world models, and why software verification is more important than ever in the LLM era. A spicy fireside chat hosted by Will Wilson, CEO of Antithesis, from the lively conference Bug Bash 2026, “A conference on extracting reliable software from the slop factory”:

Bonus: Check out YouTuber Husk on using a GenAI-fueled to bargain on his behalf. LLM-powered AI agents are gonna be great! Watch to the end!

Discussion about this post

Author

is there a good set of links on the assumptions hyperscalers have med about depreciation?

May 17Edited

Great stuff, Gary -- you've been busy! (Also, great meeting you at TAIS in Oxford!)

I cover many of these issues (the fundamental cognitive inadequacy of LLM-based chatbots (and why this will never be fixed by scaling), the construction of fully-interpretable world models via continuous learning, and neurosymbolic architecture for AGI) in my paper "Gold-Standard AGI: Outer AGI Superalignment", which I have just submitted to the Springer journal AI & Ethics.

At 214 pages (including 534 references), I admit that my paper may be a little daunting to read -- considering the very high stakes (the future of all humanity for all eternity), I felt that I needed to address every sub-problem as clearly and thoroughly as possible. However, I do believe it’s possible to get at least the gist from reading the Introduction and Conclusion sections alone.

A preprint of the paper (which took me ~8 years to develop) is available here: https://doi.org/10.5281/zenodo.16876832. The next paper in the sequence (on inner AGI superalignment) is already in my head — I just hope it doesn’t take another 8 years! :-)

Ready for more?

来源:Gary Marcus:The Road to AI We Can Trust(RSS)· garymarcus.substack.com