There’s a growing consensus that we can’t LLM our way to AGI. Today’s approach of trying to train AI on every scenario it can encounter instead of intelligent computers that can learn is a callback to expert systems from the 1980s.
We can create impressive systems with this approach such as Waymo’s self driving cars but you can’t replace human workers with this approach. Not until a breakthrough in creating systems that can learn from experience.
https://www.dwarkesh.com/p/thoughts-on-ai-progress-dec-2025