Game over for pure LLMs. Even Rich Sutton has gotten off the bus (garymarcus.substack.com)

Turing Award winner Rich Sutton — author of the influential 2019 essay "The Bitter Lesson," which argued that scaling general-purpose methods would outcompete hand engineering — has publicly softened his stance, aligning with a growing chorus of AI leaders who now question "pure" large language model (LLM) scaling. The newsletter reports Sutton’s podcast remarks echoing critiques the author has been making since 2019; combined with similar turns by Yann LeCun and Demis Hassabis, the piece reads as a tipping point: the community is moving from “LLMs-only” to hybrid approaches that address the limits of next-token prediction. Why it matters: this signals a shift in research priorities and investment. Technical implications include renewed emphasis on world models, decision-making via reinforcement learning, better grounded/causal representations, neurosymbolic methods, and built-in inductive biases or “innate constraints” to improve sample efficiency, planning, and robust generalization. Sutton favors stronger RL integration, the author favors neurosymbolic and constrained architectures — but both agree prediction alone is insufficient. Practically, the summary argues we can and should reallocate a fraction of current LLM funding to experiments that combine learning, explicit models of environment dynamics, and symbolic structure to build more capable, efficient, and controllable AI systems.