Yang Likun's recent pessimism about LLMs is quite harsh, but the new narrative of robots + world models is indeed much more interesting than simply scaling up parameters. Whether AMI's bet will pay off remains to be seen in 2027.

In the unsupervised learning podcast, Yang Likun presents a non-mainstream view on the limitations of large language models, explores the path of robotics technology, explains the reasons for leaving Meta, and diverges from Hinton and Bengio on the Turing Award perspective. He also predicts 2027, introduces the new company AMI's bet on world models, compares OpenAI and Anthropic to Sun Microsystems; and suggests that PhD students stop researching LLMs, sharing insights on AI safety, breakthrough research mechanisms, and the gains and losses of Meta FAIR.