The conversation critiques the dominance of Transformers, suggesting the field is trapped in a local minimum where scaling masks fundamental flaws in representation and reasoning. Using the 'spiral' analogy, the speakers illustrate how models mimic patterns without understanding them. They introduce the Continuous Thought Machine (CTM), a biologically inspired architecture that enables step-by-step reasoning, adaptive computation time, and natural backtracking—behaviors absent in standard LLMs. Unlike fixed-step models, CTM allocates thinking time dynamically, mimicking human-like problem solving. The model’s design supports uncertainty awareness and iterative correction, demonstrated through maze-solving and Sudoku Bench, a new benchmark emphasizing meta-reasoning and real-time strategy adaptation. The discussion advocates for research freedom and long-term exploration over benchmark chasing, positioning CTM as a path toward more authentic machine cognition.