Escaping the LLM bubble: AMI Labs goes the physical AI route

This article translates and summarizes an episode of Zhang Xiaojun’s podcast series featuring Saining Xie, co-founder and chief science officer at AMI Labs.

Saining Xie, co-founder and chief science officer at AI startup AMI Labs, warns that the AI industry is making a mistake by obsessing over large language models (LLMs). He believes that by relying on language to train machines, the industry is distracting researchers from the harder challenge of building systems that actually understand the physical world.

Rather than feeding machines more text, leaders must acknowledge the limitations of language-based reasoning, according to Xie. He also points out that researchers who abandon the hype of text-based impact to focus on how machines perceive physical reality are the ones who will ultimately build independent AI.

This philosophy directly drives AMI Labs, which Xie co-founded alongside Turing Award winner Yann LeCun. In March, the company closed a historic US$1.03 billion seed round to build advanced “world models.” These AI systems learn from raw sensory data to safely power robotics, healthcare, and industrial automation.

Rejecting how researchers are measured

The pressure to publish papers and gain industry recognition frequently warps the goals of scientific research. In the modern tech landscape, appearing to make progress has become more important than making actual discoveries.

This focus on changing the world ignores ethical considerations. Xie notes that while “creating impact is fine in itself,” it remains a self-centered pursuit. 

“I am going to change the world, but do the people in this world agree to be changed by me?” he then asks.

Seeking understanding over influence
For Xie, the purpose of research is more fundamental. 

“The purpose of doing these things is not to create impact but for understanding itself,” he explains.

Xie also believes that grasping a difficult concept brings satisfaction. The ultimate reward comes from sharing that clarity.

“If you can write down what you have understood and spread it, then you can potentially allow more people in the world to understand a question in the same way you do,” he further notes.

Real discovery takes a lifetime
This patient approach to learning directly contradicts the current pace of the tech industry. It emphasizes that breakthroughs don’t happen overnight.

“This is not a momentary burst of hormones or adrenaline,” Xie says, adding that he sees discovery as a quiet process built over a lifetime.

Embracing the long-term view of discovery

How scientific discovery actually works

The limits of language

Building systems that understand space

Entrepreneurship outside the tech bubble

---

---