AI Research Legend’s Honest Assessment of Where We Are

Transformer architecture co-author Lukasz Kaiser examines the current state of AI, questioning whether reasoning and next-token prediction are sufficient for true generalization. While current models demonstrate impressive capabilities in coding and mathematics through reinforcement learning, they remain data-hungry and prone to "sharp edges" that necessitate constant human oversight. The emergence of coding agents has transformed research productivity, enabling rapid experimentation and the testing of non-traditional architectures on accessible hardware. Despite the dominance of large-scale models, the vast space of potential learning algorithms remains largely unexplored. Future breakthroughs may require moving beyond current paradigms to mimic the human ability to learn concepts from minimal data. The competition between closed-source labs and the growing open-source ecosystem continues to drive innovation, though the path toward more efficient, autonomous research assistants remains a significant, unresolved challenge.