As large language models (LLMs) dominate the artificial intelligence market, a prominent voice has argued they still possess clear limitations in understanding the physical world. Simultaneously, new research showing that a human-like reasoning structure spontaneously formed inside an LLM has reignited the debate over artificial general intelligence (AGI).

Pascal Fung, co-founder and Chief Research & Innovation Officer (CRIO) of AMI Labs, delivered a keynote at the International Conference on Machine Learning (ICML) 2026 at COEX in Seoul on July 7. “LLMs only understand the world indirectly through text written by humans,” Fung said. “For AI agents that operate in the real world, we need a world model that directly understands the physical environment.”

Fung used a soccer match as an analogy. A player instantly reads the spatial relationships between the ball and opponents, the causal links between physical actions and outcomes, and even grasps team objectives, communication with teammates, and emotions. “A true world model reads both the physical and mental worlds together like this,” he explained.

AMI Labs is a startup founded late last year by Yann LeCun, a professor at New York University, former chief AI scientist at Meta, and a globally recognized AI authority. The company drew industry attention by raising $1 billion (~1.5 trillion won) in seed funding this year alone. LeCun has consistently argued that LLMs alone face clear limits in extending AI into the real world.

Fung pointed to the learning objective as the fundamental problem with LLMs. “The direction of LLMs aims for fluent conversation and linguistic expression, not understanding the world,” he said. “They are ill-suited to grasping the physical causality of the real world.” He added, “Because LLMs operate by generating the next token through massive computation, they are much slower and more costly than humans even when trying to understand the world.”

The hallucination problem of LLMs was also cited as an obstacle to real-world expansion. “Hallucinations occurring at the text level may be harmless, but if a hallucination occurs in a robot, it could lead to a collision,” Fung warned.

Indeed, even top-tier models like OpenAI’s GPT and Anthropic’s Claude fall far short of humans in physical world understanding. On DeepPhy, a benchmark evaluating models’ physical reasoning performance, humans achieved a 64.7% accuracy rate, while the highest score among AI models was just 41.2%.

Fung presented perception, prediction, planning, and memory as essential elements for AI agents operating in the real world. He also noted that current Vision-Language-Action (VLA) models mostly rely on imitation learning and become vulnerable when they deviate from the training data distribution, stressing that “a world model not tied to specific actions is needed.”

AMI Labs is researching world models based on technologies like JEPA (Joint Embedding Predictive Architecture), which LeCun has studied since his time at Meta. Unlike LLMs that generate the next pixel, JEPA focuses on understanding the world by predicting the next situation. This path differs from video-generation AI extensions like Nvidia’s Cosmos or Google DeepMind’s Genie.

“Physical AI is an area where many things remain unsolved,” Fung said. “We need robotic agents that can learn in a zero-shot manner even in environments with far less data.”

Anthropic Finds Traces of AGI Inside Claude

At the very moment LeCun was emphasizing the limits of LLMs, Anthropic announced research findings that it had discovered evidence approaching AGI inside its own Claude model.

According to a paper published by Anthropic, “Global Workspace in Language Models,” a special region called ‘J-space’ exists inside Claude. This region controls Claude’s thinking and reasoning; without J-space, Claude reportedly drops to a level where it can barely answer kindergarten-level questions.

Notably, J-space is not something Anthropic intentionally designed but a structure that Claude spontaneously formed over its long training process. The Anthropic research team was ultimately unable to determine exactly how J-space emerged.

Drawing an analogy to human brain activity, most tasks—breathing, posture adjustment, etc.—are processed unconsciously, while conscious thought occurs intensively in specific regions. Anthropic explained that a similar division appears inside Claude. Most tasks are processed automatically, but when complex reasoning is required, information concentrates in J-space, a ‘concept space.’

Using an observation tool called the ‘Jacobian lens’ (J-lens) developed by the researchers to peer into J-space’s workings, it was revealed that when Claude answers a question, it first recalls key concepts in J-space and then develops its reasoning based on them. For example, when asked, “How many legs does an animal that spins webs have?” Claude first activated ‘spider’ in J-space, then answered ‘eight’ based on that.

Even more striking is that artificially altering a concept in J-space completely changes Claude’s answer. When researchers changed ‘spider’ in J-space to ‘ant,’ Claude immediately answered ‘six.’ Changing ‘France’ to ‘China’ caused it to answer questions about France with China’s capital and currency. This is evidence that J-space is not merely a record of results but directly participates in the subsequent reasoning process.

Anthropic mapped J-space’s characteristics onto the properties of ‘consciously accessible information’ discussed in human cognitive science: the content of J-space can be reported, it is actively activated on demand, a single concept is flexibly utilized across multiple tasks, and it operates selectively.

This discovery provides a significant turning point in the AGI debate. In the past, a model’s AGI status was judged by external performance such as solving math problems, coding ability, or passing exams. However, the J-space case demonstrates the possibility that models can internally and spontaneously form human-like reasoning structures.

LeCun’s AGI Skepticism

Ahead of this, LeCun took to his X (formerly Twitter) account on July 4 and bluntly stated, “The ‘G’ in AGI is nonsense.” Since 2023, he has consistently argued that “ChatGPT, Claude, and Gemini are not the path to human-level intelligence.”

LeCun’s logic is clear. Exam questions are verbalized, discretized tasks—precisely the area where LLMs excel. True intelligence must include abilities stemming from common sense, such as perception, physical intuition, and causal understanding.

In an interview, he offered the following example: “What happens if you stand a pen upright and let go? Even a young child knows the pen will fall. But no one can predict which direction it will fall. An LLM will generate a plausible prediction based on statistical patterns, but because it is performing statistical completion rather than reasoning about physical reality, it will almost certainly be wrong.”

LeCun argues that AI agents must observe the world, learn how the world changes, and then internally formulate predictions and plans before acting. LLMs have learned vast linguistic knowledge, but this is merely an interface for interaction, not an understanding of the world itself.

He also opposes the Silicon Valley belief that simply scaling up model size will cause AGI to emerge spontaneously. His reasoning is that language data alone cannot fully capture the real world, and pure autoregressive prediction is not an efficient method of reasoning.

However, what LeCun opposes is less the concept of AGI itself than the Silicon Valley tendency to overhype current AI systems as imminent AGI and emphasize doomsday risks. He warns that this attitude inflates excessive expectations for AI and could ultimately cause the AI bubble to completely collapse.

Alongside CRIO Fung, AMI Labs Chief Scientific Officer (CSO) Xining Xie and others were also reported to have visited South Korea for ICML 2026. During the conference, AMI Labs plans to hold a joint event with SBVA (formerly SoftBank Ventures) to share its research direction with global researchers.