If It Only Had a Brain: The Emergence of Artificial General Intelligence

Since its launch in November 2022, ChatGPT has been dubbed everything from “mind-blowing” (CNET) to the artificial intelligence (AI) industry’s “next big disrupter” (The New York Times) – and with good reason.

The brainchild of OpenAI, ChatGPT applies machine-learning algorithms to OpenAI’s GPT-3.5 family of large language models, fine-tuned with both supervised and reinforcement learning techniques, to interact conversationally with users in a way that sounds surprisingly humanlike. Using previous questions and answers to inform its next responses, ChatGPT is able to remember a dialogue thread in a way that sounds both creative and authoritative. It can also create content, generate AI art, write code and convert unstructured data into a structured format.

Although not without its limitations – OpenAI readily admits that it sometimes provides plausible-sounding but incorrect answers and can often be excessively verbose – ChatGPT has already attracted a million users and led many experts to conclude that it poses a serious threat to Google. With the global chatbot market expected to exceed $10.5 billion by 2026, ChatGPT clearly represents not only an incredibly powerful tool but a major leap forward in self-learning artificial intelligence.

By its own admission (see sidebar), ChatGPT is not artificial general intelligence (AGI). Although its capabilities are most impressive, ChatGPT is really manipulating symbols without any understanding of what those symbols mean. And until that is possible – until today’s artificial intelligence advances to something that has the ability to actually understand any intellectual task that a human being can – AGI remains a hypothetical (although many would argue inevitable) possibility.

But we can look at ChatGPT’s capabilities and limitations to conclude a number of interesting and useful things about the future of AI and the emergence of AGI.

1. Many shortcomings of today’s AI that are impeding it from becoming AGI can be overcome.
2. AGI’s emergence will be a gradual process rather than a “singularity.”
3. It will become progressively difficult to determine whether some future system is an AGI.

What’s Needed for AGI

To become an AGI, a system will need to add the ability to comprehend that physical objects exist in a physical environment – words are most useful if they represent objects, thoughts or ideas that exist in a reality. ChatGPT’s words exist only in the context of other words, not the abstract concepts they might represent. Further, there are many concepts that are difficult to express with words alone. Consider the difficulty of describing a work of art or music in words without a picture or recorded sound.

This limitation can be overcome by adding multisensory inputs to an AI and, more importantly, an underlying data structure that supports creating relationships between multiple types of data. Physical objects have not only visible shapes and colors but textures and other tactile sensations, aroma, taste, known interactions with other objects, and a characteristic of persistence. Consider your understanding of a soap bubble. The words are just a shortcut for your multisensory understanding of its appearance, ephemeral nature, interaction with other bubbles and other physical objects, and a host of other sensations.

To capture something akin to your understanding of the bubble, an AGI will require an internal mental model of its surroundings with the entity at its center, resembling the way in which humans view the world around us from our own point of view. How could a system understand or mimic the concept of a “point of view” without ever having experienced one? An artificial entity must also possess a perception of time, which allows for an understanding of future outcomes based on current actions.

Finally, AGI will only be possible if the artificial entity has an imagination, which allows it to consider multiple possible actions, evaluate their potential outcomes and then choose the option that seems most plausible. Is this a complete list of capabilities needed for AGI? Of course not. But the implementation of these generic capabilities will lead us to a better understanding of what remains to be addressed on the road to AGI.

These AGI concepts are not impossible to create but require a mind shift in our approach to AI. Starting with a text-based system like ChatGPT and attempting to extend it to handle multisensory information, a mental model, cause and effect, and the passage of time is a daunting prospect. But starting with a data structure and set of algorithms to handle this diverse information and extending them to the text abilities of GPT with the vision, planning and decision-making of an autonomous robot is a much more promising approach.

AGI vs. the Human Brain

When considering the likelihood of such a system, understand that the human brain does all these things in an organ that weighs about 3.3 pounds and uses about 12 watts of energy. We know a lot about the brain’s structure and the way its neurons work but have yet to fill in the complete picture. Biological neurons function differently from the “perceptrons” of today’s neural networks at a basic level, and this accounts, in part, for the unparalleled energy efficiency of the brain.

We do know that the initial structure of the brain is defined by our DNA, and the complete human genome is only about 750 MB. What fraction of our DNA defines the brain? We don’t know because the structure of any human part is largely dependent on its chemistry, and we have a much better idea of how DNA applies to chemistry than about how it applies to structure. But somehow, the DNA defines the structure of a person, which includes the structure and chemistry of the brain. Once the brain is “turned on,” its structure immediately begins to modify itself to handle the data and experiences it receives. We can presume, though, that generalized intelligence is a direct outgrowth of the structure that is defined by our DNA.

The neocortex is the part of the brain humans use to think. Maybe its structure is defined by only 1% of that DNA, maybe 10%. Again, we really don’t know, but that means we might be able to define a complete AGI in a program that is as small as 7.5 MB. And 7.5 MB is a program that is well within the scope of a research team to write in just a few years. Thus, the AGI problem is really what is going to be written, not what the underlying structure requires gigabytes to define. At this point, we simply don’t know what to write as the fundamental AGI algorithm. The massive computer programs of today’s AI might not even be necessary for AGI.

AGI Emergence Will Be Gradual

The emergence of AGI is likely to be gradual rather than all at once because advances in AGI development will create solutions or capabilities that will be individually marketable. Consider how adding a generic mental model would enhance any mobile system, such as an autonomous robot or self-driving car. Consider how improving language skills has already revolutionized personal assistants and online search. Consider how understanding the interactions of things will improve system decision-making.

All these capabilities already exist at some level. As they are merged and each advances at its own rate, systems will approach or exceed human-level ability in some areas and others will lag behind. Today’s AI has superhuman capabilities in some areas and can’t match the abilities of most 3-year-olds in others. This situation will continue, and AI will advance sporadically in different areas – who is to say a system that (for example) can pass graduate-level exams in many areas but can’t make a cup of coffee is a “general intelligence” or not?

Impeding long-term AGI development is the commercial need for short-term results. Take, for example, the current state-of-the-art computer vision used in self-driving cars, which is generally inadequate for safety. To solve this problem, the vision of self-driving cars is augmented with other sensors (radar, lidar, ultrasound, etc.) that provide more absolute information about the car’s surroundings. Even though this represents a great solution to optical vision in self-driving cars, it may or may not contribute to advancing an overall AGI system. The additional sensors, in fact, may “solve” the present problem in a way that relegates improvements to the underlying vision system to the back burner.

We Won’t Know When AGI Has Arrived

Humans have widely varying levels of ability. The Turing test has long been an exemplar for general intelligence, but as systems like ChatGPT get closer to human-level language abilities, the goalposts need to be moved. It is clear that for ChatGPT to pass a Turing test, it would have to be reprogrammed to lie about its own abilities and limitations – not a desirable outcome.

Given that, a few alternatives have been suggested, such as the “coffee test,” “robot college test” and “flat pack furniture test.” Taken together, these attempts at tests show our limited understanding of the nature of general intelligence itself. Furthermore, there are many people who could not pass various tests because of physical or cultural differences but otherwise seem to be generally intelligent. And none of these tests seem to address some questions that most children are good at, such as “what will happen next?”

But even with these tests, where would we place a system that can pass the robot college test but is too clumsy to make a cup of coffee? This means that we will continue to argue about whether AGI has been achieved for perhaps a decade after credible experts begin saying that it has been reached. By that time, systems will have surpassed human abilities in so many areas that we will be forced to agree that some sort of AGI has been created.

AGI is Inevitable and Sooner Than You Think

As we learn more and more about brain function, the only argument against the eventual creation of AGI is that the human mind contains some fundamental “magic” that is outside the scope of physical systems, which we can replicate in nonbiological systems. That position aside, if all of today’s more specialized, individually marketable AI systems were built on a common underlying data structure, they could begin to interact with each other, building a broader context that can actually understand and learn. As these systems become more advanced, they will be able to work together to create a more general intelligence that approaches the threshold for human-level intelligence, then equals it, and eventually surpasses it. Perhaps only then will we humans begin to acknowledge that yes, maybe AGI does exist. It is inevitable because market forces eventually will prevail. All we are waiting for are the insights needed to make it work.

When will that happen? Some experts say it will take hundreds of years before we know how the brain works, what its algorithms are and what its capacity is, but I would suggest that it is only a matter of time before we shift from our current dependence on bigger and bigger data sets and recognize that to approach true understanding, we must address the brain knowledge gap.