Amidst the AI sector’s rush to label their developments as “AGI” or “superintelligence,” Alexandre LeBrun, the CEO of AMI Labs, a venture under Yann LeCun’s world model initiative, does not employ these terms at all. LeBrun shared during a conversation with TechCrunch that their company completely steers clear of phrases like “AGI” or “superintelligence.”

“We’ve never used the term AGI. And I’ve observed that nobody seems to be using it anymore; they’ve turned to superintelligence,” he mentioned. “Next time we might adopt yet another term.” He remains skeptical about the new terminology too. “There’s no strong definition. What does superintelligence even mean? I’m uncertain. It’s not particularly helpful.”

This is a deliberate position from a founder positioned at the forefront of AI’s latest competition.

LeBrun spoke with TechCrunch during his visit to Seoul last week for The International Conference on Machine Learning, where he was searching for local industrial partners, global enterprises, and researchers. Although AMI Labs has not yet released a product, it is already engaging with sectors like robotics, manufacturing, and electronics. LeBrun clarified that a world model, which integrates physics to anticipate and interact with reality, must demonstrate its capabilities beyond laboratory settings.

One domain where world models are anticipated to significantly impact is robotics. Currently, robots merely execute fixed sequences, remaining “entirely static,” and AI still seems “quite limited in physical contexts,” LeBrun remarked.

Even the capability for AI to at least make robots “aware of their environment” would represent “a substantial change for the world.” Such contextually aware AI could have, for instance, prevented a robot performing dance and kung fu at an event from approaching and injuring a child. “The hardware has progressed immensely; the advancements in hardware over recent months are astounding, yet there’s no intelligence.”

A large language model (LLM) predicts subsequent words or text, while a world model forecasts the next state. If a glass is nudged off a table, you know it will tilt and spill; that’s the insight a world model aims to encapsulate: foreseeing the subsequent condition of reality, LeBrun elucidated.

He does not assert that world models surpass LLMs, which are “complementary, not substitutable” within AI frameworks that comprehend the physical realm, LeBrun stated. Drawing a comparison to the human brain’s separate language and reasoning functionalities, he mentioned that LLMs will continue to be the most effective tools for language processing while world models will offer context and real-world insights.

Virtually every industry that interacts with the “real world” could ultimately harness robotics founded on world models, LeBrun asserted, contending that physical environments are still where LLMs tend to falter.

A factory robot performing repetitive actions functions adequately today, he noted. The challenges arise when “you introduce your robot to an open environment, whether at home or on the street,” where it needs to comprehend its surroundings and act safely. “Currently, robots are not safe,” he stated. “There’s no solution to that at present.”

Healthcare provides a more personal illustration for LeBrun, whose former company was Nabla, an AI health startup. He compared the current AI systems to a doctor trained solely in textbooks without any practical experience. LLMs may have utility in healthcare, he pointed out, but they only address “1% of the healthcare landscape.” The remaining 99% relies on real-world experience. 

However, according to LeBrun, a world model cannot be developed in a laboratory setting. To train on practical realities, AMI requires genuine environments and committed partners, as per the CEO. “We need access to the real world,” and forming partnerships is “more efficient for us.” This is part of what draws him to Asia, where the necessary robots, chips, and factories reside.

LeBrun is not ready to outline a comprehensive strategy for Asia just yet. “It’s premature,” he remarked. However, the attraction to South Korea is rooted in two main factors. First, Korea boasts advanced industries in robotics, semiconductors, and manufacturing—the sectors that the initial phase of AI barely engaged.

The second draw is speed. LeBrun highlighted Korea’s national initiative to invest in AI and its history of early adoption. “Korea was the swiftest adopter of the internet 25 years ago,” he recalled. This combination of a solid industrial foundation alongside a propensity to rapidly embrace AI is what he considers “distinctive,” and the reason “we aim to establish ourselves here from the outset.”

“I’ve been advising Alex and the team to visit Korea,” JP Lee, the CEO of SBVA and one of AMI’s investors in Asia, conveyed to TechCrunch.

The government has been “remarkably effective” in sponsoring local sovereign LLM models, Lee remarked, and those already function “adequately” for general applications, but he is advocating for Korea to continue investing in physical AI as well. He referred to Seoul’s June plan to allocate around $880 billion for chips, AI data centers, and physical AI, as one of its three designated pillars: “They ought to coexist.”

Lee argued that Korea’s significance to foreign enterprises lies not just in hardware. Local developers rapidly adopt and tailor new tools, a phenomenon that has fostered homegrown internet entities like Naver and Kakao.

Despite its immense backing and financial support, AMI has yet to offer any products. The startup, co-founded by Turing Award laureate Yann LeCun after his tenure at Meta, secured $1.03 billion in March at a pre-money valuation of $3.5 billion. There is no product available yet, nor any timeline he is prepared to confirm. “We’ll make an announcement when we’re ready,” LeBrun stated.