AI is surpassing the LLM era; researchers point out that the next generation of AI must understand the real world.

LLM-era AI is beginning to face limitations and may not yet be the path to human-level intelligence. World-renowned AI researchers point out that the world is entering a new era of "World Models" that understand reality like humans.

July 4, 2569 - For years, the world has been thrilled by the capabilities of AI models like ChatGPT, Claude, and Gemini, leading many to believe these technologies are bringing us closer to artificial intelligence comparable to humans. However, one pioneer in the AI ​​field views this as only the beginning. He believes current AI cannot truly "understand" the real world, and that to create AI that can think, analyze, and perform real-world tasks like humans, industry needs to move beyond Language Learning Models (LLMs) to a new generation of AI architecture, which is becoming the next battleground in the global technology market.

According to a BBC article citing Yann LeCun, one of the most important figures in the field of AI, "Currently, we don't have a robot that understands the physical world as well as a mouse."

Previously, Yann LeCun worked at Meta, Facebook's parent company, as a lead AI scientist before leaving in 2025 to found a new company called Advanced Machine Intelligence Labs. His goal is to develop AI that surpasses the limitations of current systems, such as ChatGPT, Claude, and Gemini.

Yann LeCun, founder of Advanced Machine Intelligence Labs. Image from Facebook. Yann LeCun

LeCun gave an interview to reporters during the event. VivaTech At France's largest technology conference, it was stated that while these AI systems are useful in many areas, they cannot handle complex real-world situations, such as robots performing household chores instead of humans. These systems are not a path to human-level intelligence, or even animal-level intelligence, because they cannot process information from the real world. These systems were not designed for that.

This is why AMI Labs, based in Paris, is trying to develop a new type of AI that doesn't rely on the same technology as ChatGPT or other competing models at this time.

Investors also view this approach as having great potential. Earlier this year, AMI Labs announced it had raised significantly more funding. $1 billion USD (approximately £760 million) The investors include: Nvidia A major U.S. computer chip manufacturer and a private wealth management fund. Jeff Bezos The Amazon founders' Seed funding round, the very first round of funding for a startup, became one of the largest Seed funding rounds in European history.

LeCun explained that LLM models like ChatGPT have outstanding capabilities in many areas, including programming, solving mathematical problems, and generating messages. However, he views these tasks as problems with clearly defined boundaries and predictable outcomes.

"LLMs are simply systems that accumulate a collection of knowledge. AI uses a method of retrieving previously learned data. You train AI to recount what it has memorized, but that doesn't mean the AI ​​is exceptionally intelligent, because AI doesn't have a fundamental understanding of what it's doing."

LeCun believes that the real world is full of a vast number of possible outcomes for every action, therefore a new form of AI that is more flexible than current systems is needed. During an interview, he picked up a pen, held it vertically, and then asked:

"What happens when you let go?"? " Everyone knows that the pen will fall, but no one can say which way it will fall because it's impossible to predict. Conversely, an LLM (Low-Level Marketing) might attempt to construct a single answer about the direction the pen will fall, relying on statistical patterns from learned data, and that answer is almost certainly wrong.

The reason is not that the system is using reasoning to understand the physical reality of the situation, but rather that it simply generates an answer that “seems statistically likely.”

LeCun stated that the system his company is developing is called... Joint Embedding Predictive Architecture (JEPA) Designed specifically to address these types of problems, the JEPA system works by creating an "abstract image" of the real world, allowing the AI ​​to efficiently assess the potential outcomes of various actions.

Creating these abstract images requires highly complex, advanced mathematics, but in principle, the system filters out unnecessary information, leaving only the essential data that helps the AI ​​effectively understand the world.

In the case of a pen resting on its tip, the AI ​​would immediately understand that trying to guess which way the pen would fall is pointless, as there is no data available to definitively determine that direction. Creating such flexible AI is a key goal for the global robotics industry.

The BBC article states that billions of dollars have been invested in developing humanoid robots over the years, and their capabilities have improved impressively each year. However, training robots to safely perform household chores, such as ironing or arranging dishes in the dishwasher, remains more difficult, costly, and time-consuming than many realize.

While LeCun believes that current AI models are unlikely to perform well in this type of environment.

"LLMs are almost hopeless when it comes to robotics. The claim that simply increasing the size of LLMs will allow us to create intelligence superior to humans is, in my opinion, unlikely to happen in the long run."

Many researchers in the AI ​​industry agree with LeCun's viewpoint; one of them is: Ingmar Posner Professor of Applied Artificial Intelligence จาก University of Oxford He serves as the director of the university's Applied AI Lab and also functions as an Amazon Scholar.

Posner stated, "In my view, the next decade will be the era of AI systems that can explain reasoning," adding that the world needs models that can answer important questions such as:

• What is really important?
• What causes what?
• If we chose a different approach, how would the outcome change?

Posner and a team of about 10 researchers spent over four years developing a new type of AI, which belongs to a group of technologies collectively known as “AI (Artificial Intelligence, Artificial Intelligence, and AI).”World Models"

Although the concept of World Models has been around for decades, one of the key inspirations for modern research comes from a highly influential study published in 2018 by... David Ha Jurgen Schmidhuber

The core of this concept is that as machine learning technology and computing power advance sufficiently, AI will be able to learn how to do things by “simulating its own imaginary world” without needing to experiment in the real world every time. Since 2018, this concept has become a major driving force behind a large amount of research on world models.

One of the most striking examples is: Dreamer World Model ของ Google Last year, one version of Dreamer was able to learn how to collect diamonds in the game. Minecraft This can be achieved by imagining various possible scenarios in advance, before using that information to inform decision-making.

Posner hopes that the AI ​​system his team is developing will be another significant step in this field of technology, which he calls... Mechanistic World Model This will serve to organize knowledge in a way that AI can efficiently reuse it. However, it is currently very difficult to estimate how long it will take for these new forms of AI to be fully developed.

Posner gave an example: "If you went back to 2017 or 2018 and asked how long it would take to have a system like ChatGPT, the answer you would get was decades, even many decades. But in reality..." The first version of ChatGPT. It was officially launched in November 2022, much sooner than most experts had predicted.

The BBC article states that the development of World Models technology did not occur solely in Posner's research lab; currently... Google DeepMind which is a subsidiary of A We are currently developing a model called... Genie Meanwhile, an AI startup from London... Wayve It also has its own system called... Gaia

side Fei Fei Li One of the world's pioneers in AI has founded a company. World Labs It was established in San Francisco in 2023 to develop next-generation AI models as well.

Regarding LeCun's AMI Labs, he stated that for the remainder of this year, the company will focus on improving and further developing its AI models to make them more efficient. Then, next year, the company aims to begin deploying these AI models in real-world applications, starting with the industrial sector. If the trials are successful, it will then be time to expand the scope of their applications.

LeCun stated, "Ultimately, we will have general generic intelligence systems that can be applied to almost anything in the world, requiring minimal additional training or model refinement."

When asked what role humans would play in a world where robots can work independently,

LeCun responded, “We still need humans to ask the questions: what should we ask, what should we create, and what should we invent. That is what truly defines humanity.”

He believes that AI will function as a "helper" to humans, rather than replacing them.

"The interaction between humans and AI in the future, even if one day AI becomes smarter than us, will be not unlike the relationship between organizational or political leaders and their teams of assistants, many of whom may be more capable or intelligent than the leader themselves."

LeCun believes that no matter how advanced AI becomes, humans will still be the ones setting the direction, questioning, and creating new things. AI will serve as a tool to help humans achieve their goals more efficiently, not as a means to completely replace human roles.

refer : www.bbc.com

Read news related to Technology circles around the world can be found here.