Descriptions:
In this solo essay-style video, Dwarkesh Patel lays out a pointed critique of the current AI scaling consensus, arguing that there is a fundamental logical contradiction between researchers who hold short AGI timelines and the industry’s heavy investment in reinforcement learning environments designed to pre-bake specific skills into models.
Patel’s central thesis draws on a blog post by Baron Millig: if frontier models were truly approaching human-like general learners, the billions spent teaching them to use Excel, navigate browsers, and handle consultant tasks would be unnecessary. He cites the robotics domain as a particularly vivid example — a human-like learner would make robotics largely a solved problem, yet labs are still doing massive environment-specific training runs. He references a dinner conversation with an AI researcher and a biologist to illustrate the gap between ML optimism and domain-specific reality.
The video also addresses the counterargument that pre-baking skills is simply more efficient than on-the-job learning, granting its validity for common tools while arguing it breaks down for the long tail of company- and context-specific knowledge. Patel closes by noting that if models were genuinely at AGI capability, firms would be spending trillions on tokens — and the gap between current revenue figures and that projection is itself strong evidence that the capability bar has not been cleared.
📺 Source: Dwarkesh Patel · Published December 23, 2025
🏷️ Format: Opinion Editorial
