First slowly, and then all at once, dreams of LLMs bringing us to the cusp of AGI have fallen apart. The last few months have been devastating:
• June, 2025: the Apple reasoning paper confirmed that even with “reasoning”, LLMs still can’t solve distribution shift, the core Achille’s heel in neural networks that I have been writing about for nearly 30 years. Several other papers, include the ASU “Mirage“ paper quickly extended this finding.
• August, 2025: GPT-5 came late and fell short.
• September, 2025: Turing Award winner Rich Sutton, known best for RL and his “bitter lesson” thanked me for my critiques of LLMs and agreed. Seismic. (And astonishingly gracious!)
• October, 2025: Andrej Karpathy, a widely-respected machine learning expert who once ran AI for Tesla and twice worked for OpenAI, just said agents aren’t anywhere close, and that AGI is a decade away.
• October, 2025: Nobel Laureate and Google DeepMind Sir Demis Hassabis just blew up some wildly overhyped claims from OpenAI’s Sebastien Bubeck about math. Researcher/Hyperbolic Labs CTO Yuchen Jin explains:
Embarrassing indeed.
LLMs have their place, but anyone expecting the current paradigm to be close to AGI is delusional.
Gary Marcus tried to tell you for years that this moment would come for years. Do consider reading his 2020 article Next Decade in AI, for an alternative strategy that increasingly looks prescient.
And please consider joining almost 85,000 others in subscribing to the newsletter that told you so much that has proven true, long before it was fashionable.
‘ The preceding article may include information circulated by third parties ’
‘ Some details of this article were extracted from the following source garymarcus.substack.com ’
