
Dwarkesh discusses the tension between short AI timelines and the current approach of reinforcement learning on LLMs, arguing that if AI were close to human-like learning, the current training methods would be unnecessary. He critiques the idea of pre-baking skills into models and questions the plausibility of AI researchers solving AGI without basic learning capabilities. Dwarkesh suggests that the focus should be on continual learning, where AI agents learn from real-world experiences and share knowledge, but acknowledges that achieving human-level on-the-job learning will take time. He anticipates incremental progress in continual learning, rather than a sudden breakthrough, and expects fierce competition among model companies to continue. He also reflects on the shifting goalposts in AI evaluation, noting that while models have made significant progress, they still fall short of achieving true AGI and generating substantial economic value.
Sign in to continue reading, translating and more.
Continue