Imagine you are trying to teach a robot to understand the world. For the last decade, the biggest breakthroughs in artificial intelligence (AI) have come from feeding these robots massive amounts of digital information—text, images, code, and videos. This is like giving a student an endless library of books about the ocean. They can memorize every fact about tides, fish, and storms, and they can write beautiful essays about the sea. But if you ask them to actually swim, to feel the cold water, or to navigate a current they've never seen, they might flounder. This is the current state of "Narrow AI": systems that are brilliant at processing digital data but struggle with the messy, real-world stuff like cause-and-effect, understanding human intentions, or logical reasoning that requires more than just pattern matching. Scientists call the goal of building a machine that can do all of this "Artificial General Intelligence" (AGI). The big question everyone is asking is: Why haven't we built it yet? Is it because our computers aren't big enough, or is there something missing from the way we are teaching these machines?

This paper, written by researchers Yi Peng, Huansheng Ning, and Jianguo Ding, argues that the problem isn't the size of the computer or the amount of data we have. Instead, the problem is where that data comes from. The authors suggest that current AI is stuck in a "Cyber-only" bubble, learning only from digital records. To become truly smart, they argue, AI needs to step out of the library and into the real world. They propose a new way of looking at AI learning called the "CPST" framework, which stands for Cyber-Physical-Social-Thinking. Think of it as a four-dimensional map of reality:

• Cyber: The digital world of text, code, and files (where current AI lives).
• Physical: The world of objects, gravity, and cause-and-effect (where you can push a cup and see it fall).
• Social: The world of human feelings, intentions, and rules (where you have to guess what someone else is thinking).
• Thinking: The world of strict logic, math, and structured symbols (where you can prove a theorem).

The paper suggests that to get from a "smart calculator" to a "general intelligence," an AI needs to close the loop in all four of these worlds, not just the digital one.

The Great Data Bubble

The researchers looked at eighty different AI systems, ranging from early image classifiers to the most advanced robots and chatbots available today. They sorted these systems based on how many of the four "worlds" (Cyber, Physical, Social, Thinking) they actually interact with in a "closed loop." A closed loop means the system doesn't just read data; it acts on the world, sees the result, and learns from it.

Here is what they found, and it's a bit of a shocker for the "bigger is better" crowd: No single AI system in their entire list of eighty has managed to close loops in more than two of these worlds.

The most advanced systems that are purely digital (Cyber-only) are hitting a hard ceiling. They are amazing at recognizing patterns in the data they've seen, but they fail miserably at things that require real-world understanding. For example:

• Causal Reasoning: If you show a video of a robot stacking blocks, a digital-only AI might guess the next move. But if you ask it, "What would happen if I removed the bottom block?" (a counterfactual question), it often gets it wrong. It hasn't felt gravity or pushed the blocks, so it doesn't truly understand the physics.
• Social Cognition: Current AI can mimic human conversation, but it struggles to understand what someone else is thinking if that person is hiding their true feelings. It's like playing a game of poker where you can read the cards but can't read the player's face or body language.
• Logical Reasoning: While AI can solve math problems by guessing patterns, it often fails when asked to follow a strict, step-by-step logical path that requires checking its own work against a set of rules.

The paper explicitly rules out the idea that simply making the AI bigger (adding more parameters or training it on more text) will fix these problems. They point out that even the biggest, most powerful models are still stuck in the "Cyber" zone. They are like a chef who has read every cookbook in the world but has never actually touched a stove or tasted a dish. No amount of reading will teach them how to cook a perfect meal if they never get their hands dirty.

The Missing Ingredients

The authors use a clever analogy to explain why the "Cyber-only" approach fails. They say that each of the four worlds provides a unique type of "signal" that the others simply cannot produce.

• To learn causality (cause and effect), you need Physical data. You have to actually do something and see what happens. You can't learn this just by reading a book about physics.
• To learn social intelligence (understanding people), you need Social data. You have to interact with real humans, see their reactions, and adjust your behavior. You can't learn this just by reading chat logs.
• To learn complex logic, you need Thinking data. You need a structured system of symbols and rules that you can query and modify, not just a pile of text.

The paper presents strong evidence for this. They compared AI systems that were identical in every way except for the data they were fed.

• When a robot was trained only on video of other robots moving, it couldn't plan its own moves. But when the same robot was trained on real video of itself moving and interacting with objects, it suddenly learned how to plan and manipulate things it had never seen before.
• Similarly, a math-solving AI that was just a language model got almost zero questions right. But when they connected that same language model to a strict logic engine (a "Thinking" tool), its performance skyrocketed.

The data shows a clear pattern: Systems that interact with just one extra world (like Physical or Thinking) are much more likely to reach a higher level of capability. But the "holy grail" of AGI—where a system can handle all four worlds at once—remains empty. In their survey of eighty systems, zero systems had successfully closed loops in three or more worlds. The entire region of the map where a true "General Intelligence" should live is currently a ghost town.

The Roadmap to General Intelligence

So, where do we go from here? The authors don't just point out the problem; they offer a three-step roadmap to fix it.

1. Semantic Alignment: We need to teach AI to translate between these different worlds. Right now, an AI might understand a physical sensor reading but not know how that connects to a social rule or a logical fact. We need to build bridges so the AI can say, "The robot bumped the cup (Physical), which made the human angry (Social), so I should apologize (Thinking)."
2. Unified World Models: We need to build a single "brain" that can hold all these different types of information at once. Instead of having one brain for physics, another for social rules, and another for math, we need one system that can juggle all of them simultaneously.
3. Governance: As AI starts interacting with the real world (Physical) and real people (Social), we need new safety rules. We can't just check the code; we need to make sure the AI understands the consequences of its actions in the real world.

The paper concludes with a bold prediction: The path to AGI isn't about building bigger computers or scraping more text from the internet. It's about building robots that can touch, feel, and interact with the physical world; creating systems that can genuinely converse and negotiate with humans; and designing tools that can reason with strict logic. Until AI systems can step out of the digital library and into the messy, complex, four-dimensional reality of Cyber-Physical-Social-Thinking, they will remain brilliant but limited students, unable to graduate to the level of true general intelligence.

The authors are careful to note that this is a hypothesis supported by strong evidence, but not yet a proven law. They suggest that if a purely digital AI did suddenly solve all these problems without ever touching the real world, their theory would be wrong. But so far, the evidence points in one clear direction: to build a general intelligence, we have to stop treating the world as a dataset and start treating it as a playground.