What do a software engineer, an information theorist, and a biologist have in common?

Although this question may sound like the beginning of a joke, I believe it may hold the key to a deeper connection that could allow us to unify three of the biggest scientific mysteries under a single solution. But before we do that, we must begin with the question that, for me, started it all: What is artificial general intelligence (AGI)?

What is AGI?

Most of us naturally associate intelligence with knowledge and problem-solving ability. A system that can answer questions, solve complex problems, or perform difficult calculations appears intelligent. Yet this intuition can be misleading. A computer, for example, can perform arithmetic operations millions of times faster and more accurately than any human, but we do not consider such a calculator to be AGI. The ability to perform a task, even an extremely difficult one, is not by itself what makes a system generally intelligent.

Then what really is AGI?

A few years ago, I started exploring that, and it eventually led me to write Turing Test 2.0: The General Intelligence Threshold, where I argued that the defining characteristic of human General Intelligence (GI) is not simply reasoning ability or the accumulation of knowledge. It is the ability to generate genuinely new functional capabilities — capabilities that no previous human (or other system) could perform and therefore could not have simply been acquired through training.

Human history provides overwhelming empirical evidence that such a generating mechanism exists. Humanity progressed from living in caves and wielding sticks and stones to building skyscrapers, harnessing nuclear power, decoding the human genome, and traveling into space. These achievements cannot be explained merely by transferring existing functional capabilities from one person to another. They demonstrate that humanity continually expands the set of functions it can perform.

At first, this way of thinking about AGI may sound unconventional. Yet remarkably, it is consistent with how leading AI researchers describe AGI. Not long ago, OpenAI CEO Sam Altman, while discussing AGI with physicist David Deutsch, proposed that an AI would definitely qualify as AGI if it could “…figure out quantum gravity and could tell you its story.”

Notice what makes this example compelling. It is not simply that the AI possesses vast amounts of information. A library can contain vast amounts of information without being intelligent. Rather, the AI would have expanded the set of functions it can perform by making a scientific discovery that no previous system had achieved. Therefore, the defining characteristic of a generally intelligent system (or AGI) is not that it can perform existing functions, but that it can generate new functional capabilities that were not initially part of the system and were not transferred from another system that already possessed them.

But this immediately raises a deeper question: What allows any system to acquire capabilities it did not previously possess?The more I examined this question, the more I realized that it extends far beyond artificial intelligence. In fact, the same question seems to lie at the heart of some of the deepest mysteries in science.

Three mysteries but one common question

What explains the origin of life? Can computer systems and AI achieve AGI? How does information convey meaning?

These are among the deepest questions in science, and they are usually treated as separate problems, studied by different disciplines using different tools. Biologists investigate the origin of life. Computer scientists pursue AGI. Information theorists ask how information is represented, transmitted, and measured. Because these fields study such different systems, it is natural to assume that they are asking fundamentally different questions.

But are they really different problems? A living cell is not remarkable because it contains molecules. It is remarkable because those molecules collectively perform metabolism, replication, and countless other biological functions. Likewise, a generally intelligent system is not remarkable simply because it stores information. Libraries contain vast amounts of information, yet no one would call a library intelligent. Intelligence emerges from the ability to use information to perform functions and, more importantly, to discover new ones.

Information itself presents the same puzzle. Information only becomes useful when some system can use it to perform a function.

These observations reveal a surprising connection. All three questions — origin of life, AGI and information — revolve around the same fundamental concept: the ability of a system to perform a function. A living cell, an intelligent system, and a piece of information appear very different. Yet they all share something fundamental: they enable a system to accomplish something that it otherwise could not. Perhaps these three mysteries are not independent after all. Perhaps they all point toward a common question: What is the relationship between information and function?

Searle’s Chinese Room thought experiment

To understand how information enables a system to perform a function, we must first understand the role of interpretation. John Searle’s famous Chinese Room thought experiment provides a useful way to explore this question.

Imagine a man locked inside a room. The man cannot speak Chinese. However, inside the room are instruction manuals written in a language he understands. These manuals explain how to manipulate Chinese characters according to specific rules to create responses in Chinese. People outside the room write notes in Chinese and pass them through a slot in the door. By systematically following the instructions in the manuals, the man produces responses in Chinese. He then passes them back to the people outside. To the people outside, it appears that the room contains someone who understands Chinese. This raises the following philosophical question: Does the man inside the room actually understand Chinese?

Searle (1932–2025) argued that he does not. The man is only manipulating symbols according to instructions, similar to the way that a computer executes a program. Others have argued that the entire system—the man together with the instruction manuals — does understand Chinese. But perhaps both sides of this argument are missing the critical point.

To understand what they are missing, imagine someone sends a note through the door, written in Chinese: “The key that unlocks the door is hidden beneath the third-floor tile. Pick it up and use it to exit the room.”

Would the manuals, by themselves, enable the man inside the room to escape? The answer is no. The manuals only instruct the man on how to manipulate Chinese symbols and produce responses; they provide no instructions for taking the additional actions required to exit the room.

To escape the room, the man (or any system) requires two different capabilities: First, the man or system must be able to interpret the Chinese message and understand what action it describes. But second, it must know how to perform the physical action required to escape: find the key, pick it up, and use it to unlock the door.

The man already possesses that second capability. He knows how to manipulate objects and use a key to open a door. Therefore, if a system inside the room possesses the ability to interpret the Chinese message in the way that the man does, it should be able to use this information to escape. The fact that it cannot do so reveals that the system lacks the ability to interpret the message and extract the function it describes.

The information in the note is not meaningless. It is meaningful for the person who wrote it, and for any system capable of interpreting its content and using it to perform the described function. The information is meaningful to these systems, but not to the system inside the Chinese room.

This reveals a fundamental property of information: Information is not meaningful simply because it exists or because it is improbable. It becomes meaningful only if there exists a system that can interpret it and use it to perform a function.

Functional information

We can now combine the insights we have gathered about information and formalize the relationship between functionality and information. Before doing so, however, let us ask this question one more time: Is functionality an intrinsic property of information itself?

At first glance, it seems like it should be. A recipe in a cookbook appears to contain the information required to bake a soufflé. Surely the functionality is already contained in the recipe.

But is that true? Are information and functionality the same? Imagine giving exactly the same recipe, written in English, to four different people:

The first cannot read English. For that individual, the recipe contains no functional information whatever.

The second can read English but has never cooked before. That person can explain what the recipe says, but cannot reliably bake the soufflé without assistance.

The third is an experienced home cook who follows the recipe and successfully prepares the dish.

The fourth is a world-class chef. The chef can not only bake the soufflé but also improve the recipe by recognizing opportunities to make it lighter, richer, or more consistent.

Notice what has changed. The recipe never changed. Only the interpreting system changed. The deeper a system can interpret the same information, the more functionality can be extracted.

This leads to what I believe is a more fundamental way of thinking about information. I call it functional information (FI).We can describe it as FI(I, S, F), where:

• I is the information.
• S is the system interpreting that information.
• F is the function the system is able to perform.

Functional information is information that a particular system must correctly interpret and use to perform a particular function.

Remove any one of these three components and functional information disappears. Information without an interpreting system cannot perform a function. A system without information has nothing to interpret. And without a function, there is no basis for distinguishing meaningful information from meaningless patterns of symbols.

Functional information is therefore not an intrinsic property of information itself.It is a relationship between information, the system interpreting it, and the function being performed.

One question to rule them all

Now that we have defined the relationship between information, systems, and functions, we can return to the three questions with which we began: What explains the origin of life? Can computers achieve AGI? How does information convey meaning?

These three questions could hardly look more different. Yet in each case, the same thing happens. A system acquires the ability to interpret information in a way that gives it a capability it did not previously possess.

The origin of life

Consider the simplest living cell. Its DNA contains information, but DNA alone is not enough to produce a living organism. The information must be interpreted by a system capable of reading it and using it to perform functions — such as producing proteins, regulating chemical reactions, and reproducing.

The important question, therefore, is not simply how the information encoded in DNA came into existence. It is how a system capable of interpreting the DNA information and turning it into biological function came into existencein the first place.

The origin of intelligence

Now consider AGI. A system that merely possesses a large amount of information is not necessarily intelligent. What distinguishes general intelligence is the ability to acquire capabilities that were not previously available to the system.

Imagine an AGI system observing astronomical data that no human has been able to explain. Such a system would not merely memorize those observations. It would need to interpret them, discover a previously unknown regularity, and develop a new understanding of how the universe works, for example, by developing a theory of quantum gravity as proposed by Sam Altman. The system would therefore be doing something fundamentally new: turning previously uninterpreted information into new functional information.

The origin of meaning

Finally, consider a signal transmitted through a communication channel. Suppose a receiver captures a stream of symbols that appears to be nothing more than random noise. The signal contains structure, but the receiver does not know how to filter or decode it. To the receiving system, the transmission is effectively meaningless.

Now imagine that the receiver discovers the correct way to interpret the signal. Suddenly, the same transmission can reveal a message, an image, or some other useful information. The signal itself did not change. What changed was the interpreting system, which has acquired the ability to extract functional information from the signal.

The one underlying question

Looking at these three examples side by side, we can now see the common thread. In every case, a system acquires the ability to interpret information in a new way and thereby obtain functionality that was not previously available to it. In other words, in all three cases a system gains functional information that did not previously have.

If this is correct, then we may have uncovered something much more profound than a connection between three scientific problems. We may have identified the same underlying problem at the heart of all three.A single, common question targets all three problems: What is the origin of functional information?

The implications are profound. If functional information is indeed the underlying problem shared by all three, then explaining its origin would provide a common framework for understanding the origin of life, the emergence of general intelligence, and the emergence of meaning. In other words, solving one fundamental problem could bring us closer to solving all three.

The true origin problem

Now that we have identified functional information as the common problem underlying the origin of life, intelligence, and meaning, we can ask a more practical question: What do we actually know about the origin of functional information?

One way a system can acquire functional information is through transfer from another system that already possesses it.This can happen in at least two ways.

The first is through explicit communication: A system can use a language that another system already knows how to interpret to describe a function. A programmer can write software that tells a computer what to do. An engineer can provide instructions for constructing a machine. A biologist can edit a genome by introducing DNA sequences that a cell already knows how to interpret. In each case, one system possesses functional information and communicates it to another system in a form that the receiving system can interpret.

The second is through training: A teacher can train a student to perform a task by providing explanations, examples, and feedback. The student gradually acquires the ability to interpret information and use it to perform functions that could not be performed before. The same principle applies to modern AI systems. An LLM can be trained on human-generated data and thereby acquire the ability to interpret and use information through a process guided by information that already exists in its training data.

In both cases, however, we are still dealing with transfer, not origin. One system already possesses functional information, and another system acquires the ability to use it. The process may be highly sophisticated, but that, in itself, does not explain where the functional information came from in the first place.

Could randomness provide the answer?

Randomness is often proposed as part of a step-by-step explanation for the origin of life. LLM “hallucinations” are sometimes similarly proposed as a possible pathway toward AGI. But randomness creates a different problem from the one we are trying to solve.

Randomness can produce new configurations of information. What it does not explain is how a system acquires the ability to interpret one of those configurations and use it to perform a function that it could not perform before. Randomness can generate possibilities; it does not, by itself, explain the emergence of the interpretation that makes one of those possibilities functional.

I explored this problem in more detail in “That Monkey Cannot Even Find Shakespeare,” where I argue that randomness cannot generate new functional information for a system if the capability to interpret and use that information is not already embedded in the system.

Perhaps some other step-by-step process could generate new functional information? In computer science, a precisely defined step-by-step procedure is called an algorithm. An algorithm starts with an initial state and applies a sequence of well-defined step-by-step rules to produce subsequent states. If the origin of functional information can be explained by such a process, then there should exist an algorithm (e.g., an AI model) capable of generating genuinely new functional capabilities from an initial system that did not already possess them.

My recent paper with Charalambos Rossides, On the Computability of Artificial General Intelligence, investigates precisely this possibility. The paper formally argues that no algorithm can demonstrate genuinely new functional capabilities that were not already present in the initial algorithm. To illustrate the basic intuition behind this argument, we can simply think of an algorithm as a state machine. An algorithm must eventually reach a state where it terminates and outputs its result; otherwise, it does not actually compute anything. For the algorithm to produce a particular result, there must already be a path through its states that leads to the terminating state corresponding to that result.

In other words, the algorithm must already contain the computational pathway required to compute the required function. If no such pathway exists in the initial algorithm, it cannot simply appear as the algorithm executes.The algorithm can traverse existing pathways and produce results that follow from its computational rules, but it cannot acquire a fundamentally new computational capability that is not generated by those rules.

If this result is correct, then its implication for the problem we have developed throughout this article is significant: no algorithmic process can explain the origin of functional information.And if the three problems we examined are indeed manifestations of the same underlying problem, then an algorithmic process cannot provide the missing explanation for the origin of life, the emergence of general intelligence, or the emergence of meaning from previously uninterpretable information.

We therefore arrive at an interesting boundary in what we currently understand. We know how functional information can be transferred. We can imagine random processes that generate new configurations of information. We can construct algorithms that systematically transform one state into another. But none of these processes, by themselves, explains the fundamental transition at the heart of our problem: How does a system acquire the ability to interpret information that it could not previously interpret — and thereby acquire the ability to perform new functions that did not previously exist within the system?

That, I believe, is the true origin problem.

A testable prediction

Throughout this article, we have developed some challenging ideas about information, functionality, and the true origin problem: the origin of functional information. But can these ideas be tested? If functional information is more than just a conceptual framework, it should lead to predictions that we can verify experimentally.

Let us then consider this simple experiment. Imagine an AI model with no prior training. Its weights are initialized randomly. We then train it exclusively on a textbook that teaches basic calculus. The question becomes, what do we expect the AI model to learn?For the purpose of this experiment, let us assume that the randomly initialized model contains no functional information relevant to performing calculus.

A natural first answer is that the AI model will learn calculus. But this seemingly simple answer is incomplete, at least according to the framework developed in this article. According to the theory developed here, the AI should learn to interpret the information contained in the book and use it in ways that the book makes possible. If our AI model contained no functional information before training, then the functional information it acquires through training must ultimately come from the information provided in the book. So, what functional information does the book actually contain?

Can the book perform calculus? Obviously not. The book can describe mathematical procedures, but it cannot execute them. It uses human language to explain calculus to a system that already knows how to interpret that language. Its functional information is therefore the information required to explain calculus, not the information required to perform arbitrary mathematical functions.

This leads to a surprising prediction. The trained AI should be able to explain calculus in ways similar to those contained in the book. It may also be able to reproduce the examples presented in the book, combine concepts from different parts of the book, or even generalize some of its explanations to examples that were not explicitly included.

But according to the theory, there should be a fundamental limit. The AI should not suddenly acquire the general ability to perform mathematics that requires functional information not contained in the book. It should not, for example, discover an entirely new mathematical technique or reliably solve classes of problems whose solutions require capabilities that were never present in the training information.

The distinction is subtle but fundamental. The model may learn how to explain calculusand manipulate the information contained in the textbook. But that does not necessarily mean that it has acquired the general functional capability of doing mathematics. And this gives us a potential experimental test of the theory.

If an AI trained exclusively on a calculus textbook can learn to explain the material, but cannot reliably perform mathematical functions that require functional information absent from the book, that would be consistent with the framework developed here. But suppose the opposite happens. Suppose we start with a genuinely untrained system and provide it only with the textbook. If that system can not only learn to explain calculus, but also acquire genuinely new mathematical capabilities that were not contained in the book — perhaps even discovering new mathematical methods or solving problems requiring functional information that was nowhere present in its training data—then the theory developed in this article would face a serious challenge.

More important, if an AI can repeatedly demonstrate this capability, then we may have discovered exactly what we have been looking for throughout this article: a process by which a system generates new functional informationrather than merely transferring or rearranging existing functional information. Such a process would provide a common framework for addressing all three origin questions.

This is also why such an experiment could serve as a powerful test for detecting AGI.If the model can only reproduce or recombine the functional information contained in the textbook, then the experiment demonstrates only the transfer of functional information from one system — the textbook and its authors—to another, the AI model. If, however, the model can generate genuinely new functional information, then it has demonstrated something far more profound: the ability to acquire a new functional capability that was not transferred from another system. That is precisely the capability that, as we have argued throughout this article, defines general intelligence.