The author expresses a cautious view on AI progress, highlighting a gap between current advancements and the expectations for achieving human-like intelligence. They question the efficacy of reinforcement learning (RL) layered on large language models (LLMs), arguing that if these models were close to achieving human-like learning, the extensive training now required would be unnecessary. Instead, they suggest that the ongoing efforts to pre-bake skills into AI models indicate a lack of foundational learning capabilities essential for artificial general intelligence (AGI). A key point raised is the difference between how humans and AI learn. Human workers can quickly adapt to various tasks without needing extensive training for each specific job. The author contrasts this with the current reliance on custom training loops for AI, which limits its ability to generalize and handle tasks that require judgment and situational awareness. They also address the misconception that slow tech diffusion is the main barrier to AI integration in firms, arguing instead that the lack of capabilities in current AI models is the real issue. If AI were truly on par with human intelligence, it would integrate into companies rapidly and efficiently. The author critiques the belief that automated AI researchers will solve the challenges of creating AGI, suggesting that it's implausible for them to tackle problems that have stumped humans for decades without possessing basic learning abilities. They underscore the need for actual AGI to generate transformative economic impacts, which they expect will emerge within the next decade or two. Their perspective is shaped by a recognition of the limitations of current AI systems and the complexities of human-like learning.