Over the past three years, the AI industry has been hurtling forward in a unified direction at breakneck speed.

From GPT-4 to Claude, and from Gemini to DeepSeek, there appears to be a growing consensus across the industry: with more data, larger models, and greater computational power, we can progressively move closer to Artificial General Intelligence (AGI).

This rationale has fueled the largest wave of global tech investment in recent times. OpenAI's valuation has soared to hundreds of billions of dollars, Anthropic has emerged as a key player in the enterprise AI market, and substantial capital has flooded into the foundational model sector. Nearly all startups are fixated on how to build applications using these models, while few are revisiting a more fundamental question—is today's large model approach truly the definitive path to AGI?

Just as the entire industry continues to pursue larger models, a scientist hailed as a founder of modern AI has chosen to start afresh.

He is Yann LeCun, the 2018 Turing Award laureate, Meta's Chief AI Scientist, a professor at New York University, and one of the three pioneers of deep learning. His newly established company, AMI (Advanced Machine Intelligence), is endeavoring to tackle a problem that starkly diverges from the industry's current mainstream trajectory.

01 ┃ What is LeCun pondering while everyone is chasing GPT?

Over the past three years, the AI landscape has been dominated by a singular logic: bigger models, more parameters, more training data, stronger reasoning capabilities...

The entire industry seems to implicitly believe that by continuously scaling up, we can gradually inch closer to AGI. OpenAI is pursuing this path, Anthropic is doing the same, Google DeepMind is following suit, and even emerging contenders like xAI, Mistral, and DeepSeek are adopting similar strategies.

But Yann LeCun has consistently stood on the opposing side. In fact, over the past few years, he has been one of the most vocal top scientists to publicly question the large language model (LLM) approach. He repeatedly emphasizes one point:

Large language models are powerful, but they don't truly comprehend the world.

This statement may seem counterintuitive—after all, today's GPT can write code, compose essays, conduct analysis, and even pass various exams. However, LeCun contends that these capabilities stem more from statistical patterns in massive datasets rather than an understanding of the real world itself. In other words, the model knows which word is most likely to follow but doesn't grasp why.

It can describe the world but doesn't truly understand it.

02 ┃ Something a three-year-old can do, but AI can't

LeCun often cites a very straightforward example: place a ball at the edge of a table, and a three-year-old child immediately knows that if it continues to roll, the ball will fall. The child has never studied physics or read any textbooks, yet they can predict the future.

Because they understand how the world functions.

Today's most advanced large models, however, lack this ability. They possess no true physical intuition. No causal understanding. No world cognition. No long-term planning. Often, they merely guess the most likely answer based on training data.

This is why, over the past few years, LeCun has been championing another concept:

World Model.

In his view, the core of human intelligence is not language but the ability to predict the world. Language is merely a tool for expression; understanding the world is the true wellspring of intelligence.

He argues that the vast majority of information humans acquire does not come from language but from vision, spatial awareness, actions, and environmental interactions. Language is just a means to express intelligence, not intelligence itself. A truly human-like AI system must not only understand text but also comprehend how the world operates, how causal relationships form, and how its own actions will yield results.

This is also the core research direction of AMI.

03 ┃ The next revolution in AI may not hinge on bigger models

If we look back at the past decade of technological history, an intriguing pattern emerges. Nearly every major technological transformation in history has not simply been about scaling up the previous generation of technology. Google was not just a bigger Yahoo, the iPhone was not just a bigger Nokia, and ChatGPT is not just a bigger search engine. What truly alters industry landscapes are new technological paradigms.

Today, more and more researchers are beginning to realize that Transformers may not be the ultimate endpoint but rather just a stage in AI's evolution. This is why more and more top labs are revisiting:

World models;

Long-term memory;

Active learning;

Causal reasoning;

Physical world modeling;

Embodied intelligence...

And AMI was born in this context.

What it truly seeks to solve is not how to generate better text but how to make AI understand the real world like humans do.

04 ┃ Why are capital markets frenziedly chasing scientists?

For investors, AMI's most alluring aspect is not its current commercial revenue.

In fact, many top investors do not evaluate such a company using the logic of traditional software firms. What truly makes it scarce is not its products or users but the possibility of creating the next technological paradigm.

Over the past two decades, Yann LeCun has helped drive the development of convolutional neural networks, which later became the infrastructure of computer vision. Subsequently, the deep learning revolution reshaped the entire AI industry.

Thus, when a scientist like him decides to launch a new venture, capital focuses not on how much revenue the company can generate this year but on whether it has the potential to influence the technological roadmap for the next decade. From this perspective, AMI represents a long-term bet on the future.

05 ┃ How vast is the market AMI is truly vying for?

Many people, upon seeing AMI, instinctively assume it is just another large model company.

But in reality, the market it targets extends far beyond chatbots. Once world models truly mature, they will impact not just the software industry but the entire real world.

Autonomous driving requires understanding physical environments.

Robots need to comprehend spatial relationships.

Industrial automation requires predicting complex systems.

Medical AI needs to understand causal relationships.

Scientific research requires building world models.

Defense systems need to simulate real environments.

Virtually all scenarios involving real-world interaction could benefit from developments in this direction. Together, these scenarios constitute a future AI economy worth tens of trillions of dollars. From this perspective, what AMI is competing for is not a specific application niche but the standard for future AI's foundational cognitive architecture.

06 ┃ Why are Silicon Valley investors so enthusiastic?

If we view OpenAI as the Microsoft of the AI era and Anthropic as the Oracle of the AI era, many investors want to know: Who will become the Bell Labs of the AI era?

Who will define the technological direction for the next decade? Who will create the next true architectural revolution?

AMI's appeal lies precisely here: it represents not a product innovation but a paradigm innovation. And history has proven time and again that the companies that create the greatest value are often not the first to build products but those who define the rules.

Conclusion ┃ After GPT, AI Has a Second Act

Over the past three years, the AI narrative has revolved around one question: Who can train the most powerful large model?

Over the next decade, the industry may begin to revolve around another question:

Who can make AI truly understand the world?

Who can build the next-generation cognitive system?

Who can define the technological path to AGI?

Yann LeCun's founding of AMI may be an attempt to answer these questions.

For most investors, the greatest value of this company today lies not in its revenue, products, or even valuation but in the new possibility it represents: while the entire industry is optimizing GPT, someone is trying to redefine AI.

Beta Fund is an early-stage investment institution deeply rooted in Silicon Valley, focusing on the most groundbreaking innovation directions in the AI era, including Agentic AI, AI Native Systems, and AI-driven vertical applications.

We continuously monitor the world's most cutting-edge AI and tech investment opportunities and have secured exclusive investment allocations for the next funding rounds of multiple star projects, including:

SpaceX

Anthropic

Netlify

Prometheus Biosciences

Anduril Industries

Figure AI

...

Author's Note: Personal views, for reference only.