Alex Irpan: To some extent, LLM is the "anti-Bitter Lesson," and world models are the future

According to 1M AI News's monitoring, Xesai Ning, Co-founder and Chief Scientist of AMI Labs, a world-modeling company that has completed a $1.03 billion seed funding round, and co-author of the Diffusion Transformer (DiT) paper, recently gave his first extensive public interview. Here is a summary of the content:

Xesai Ning systematically criticized the current AI trajectory centered around large language models, arguing that LLMs are not a true embodiment of Richard Sutton's "Bitter Lesson," and are even "to some extent anti-Bitter Lesson." He pointed out that language itself is a highly refined structure of human civilization, akin to a "shortcut," and overreliance on it will limit AI's learning about the real world.

He distinguished world models from language models: while language models predict the "next token," world models predict the "next state" based on actions, enabling them to anticipate the consequences of actions during reasoning, thus achieving a safer and more controllable form of intelligence. He also stated that "AGI is a false proposition," highlighting that creating an intelligence that can survive in the real world is more challenging than excelling in math competitions or coding.

Xesai Ning also revealed that he had twice declined Ilya Sutskever's invitations: in 2018, he chose to forgo joining OpenAI in favor of Meta FAIR, and in 2024, he politely turned down an invitation to SSI, Sutskever's new company, citing differences in their approaches to multimodality and vision.

He mentioned that AMI Labs deliberately chose not to establish offices in Silicon Valley because "Silicon Valley is already deeply entrenched in the LLM path." The company is headquartered in Paris and collects real-world data through global partnerships. He believes that the AI training paradigm is shifting from "downloading the internet" to "downloading humans," with the short-term implementation scenario being AI smart glasses and robots.