Slowly at first, then all at once, the dreams of LLM taking us to the dawn of AGI came crashing down. The last few months have been devastating:
• June 2025: Apple’s reasoning document confirmed that even with “reasoning”, LLMs still cannot resolve the distribution shiftthe central Achilles heel of neural networks that I have been writing about for almost 30 years. Several other newspapers, including the ASU”MirageThe newspaper quickly extended this conclusion.
• August 2025: GPT-5 arrived late and did not succeed.
• September 2025: Turing Award winner Rich Sutton, best known for RL and its “bitter lesson”, thanked me for my critiques of LLMs and agreed. Seismic. (And surprisingly kind!)
• October 2025: Andrej Karpathy, a well-respected machine learning expert who once led AI for Tesla and worked twice for OpenAI, just said that agents are not at all close and that AGI is a decade away.
• October 2025: Sir Demis Hassabis, Nobel Prize winner and Google DeepMind, has just blown up some wildly exaggerated claims about mathematics made by OpenAI’s Sebastien Bubeck. Hyperbolic Labs Researcher/CTO Yuchen Jin explains:
Really embarrassing.
LLMs have their place, but anyone expecting the current paradigm to be close to AGI is delusional.
Gary Marcus I tried to tell you for years that this moment would come years from now. Consider reading his 2020 article The next decade in AIfor an alternative strategy that seems increasingly prescient.
And consider joining nearly 85,000 others in subscribing to the newsletter that told you so many things that turned out to be true, long before it was fashionable.
