AMI Labs Co-founder Xie Saining: LLMs Are "Anti-Bitter Lesson" to Some Extent, World Models Are the Future

Gate News reports that on March 16, AMI Labs, the world model company that completed a $1.03 billion seed round, its co-founder and Chief Scientist, and co-author of the Diffusion Transformer (DiT) paper, Xie Saining, recently gave his first in-depth public interview. Xie systematically criticized the current AI approach centered on large language models (LLMs), stating that LLMs are not the successful embodiment of Richard Sutton’s “Bitter Lesson” and are, in fact, “to some extent, the opposite of the Bitter Lesson.” He pointed out that language itself is a highly refined structure of human civilization, serving as a kind of “shortcut,” and over-reliance on it can limit AI’s ability to learn from the real world. He distinguished between world models and language models: language models predict the “next token,” while world models predict the “next state” based on actions, enabling reasoning about the consequences of behaviors and achieving safer, more controllable intelligence. He also stated that “AGI is a false proposition,” and creating an agent capable of surviving in the real world is more difficult than solving math competitions or writing code. Xie Saining also revealed that he declined Ilya Sutskever’s invitations twice: in 2018, he chose to leave OpenAI for Meta FAIR, and in 2024, he declined an invitation to Sutskever’s new company SSI due to disagreements over multimodal and visual approaches. He mentioned that AMI Labs deliberately does not have an office in Silicon Valley because “Silicon Valley is already deeply involved in the LLM route.” The company is headquartered in Paris and collects real-world data through global partners. He believes AI training paradigms are shifting from “downloading the internet” to “downloading humans,” with short-term applications including AI smart glasses and robots.