General AI: Is it achievable?
What is General AI?
Although most people outside of the industry do not consider AI in these terms, there are actually two different types of AI that we can think about and discuss: the kind of AI we typically use in real life, for instance for doing quality control in manufacturing or for generating fashion recommendations on retail websites; and the kind of AI we tend to see in movies, which we might call “General AI.” This is AI with human-like thinking and reasoning capabilities, the kind of humanoid robots that end up overtaking the human race in science fiction films. They usually look and sound like humans, and in some tales are even mistaken for them.
But there is a reason why this General AI isn’t seen in our day-to-day lives: for experts, it seems extremely unlikely that we would be able to create it given our current approach to developing AI. Think about the “humanoid” AIs you may have seen in real life or in the news, like the Anbot in Shanghai that helps prevent riots, Josie Pepper the Munich airport attendant, and Troika the informational robot in Seoul. They all have one thing in common: they’ve been programmed to perform specific functions. If you asked Josie Pepper a question unrelated to flights and baggage claim, she wouldn’t be able to answer, because these robots are not General AI.
They don’t have human-like intelligence or the ability to “think” outside of their programming. According to one AI expert, achieving this with our current capabilities would be “like trying to take an airplane to the moon.” We can teach computers how to take on narrow skills, and we’ve now begun to combine multiple skills in them so that they can complete more complex tasks. While this certainly represents increasing Machine Intelligence, it is still not General AI. Let’s explore why, and see how we might eventually approach something like General AI. It probably won’t be anything like we imagine.
How we develop AI now
The main reason why general AI hasn’t been developed is that we have typically purposefully focused on creating AI that fulfills a very narrow, specific purpose, like the aforementioned examples of providing coordinating fashion recommendations or assessing quality in manufacturing. The more narrow the model being built, the easier it is to train and the more likely it is to achieve high accuracy. This is because most models are trained by being shown hundreds or thousands of examples and being taught to make distinctions between them or to divide them into categories. For instance, an algorithm trained to determine the type of animal in an image would be shown thousands of images of animals and slowly trained to recognize them on its own. It would be more difficult and time-consuming to train the model to distinguish between dozens of animals than it would be to train it to distinguish merely between images of cats and dogs. The former would require far more examples and therefore far more time. In AI, specificity tends to mean speed and accuracy.
It would obviously be more efficient to build a “human-like” model which, like a person, could perform a wide array of tasks, and it has been attempted many times. But so far, no approach has been successful.
GPT-3 is a recent example of an attempt at more general AI, and it certainly represents an advancement in machine learning. GPT-3 is an example of “conversational AI”: it uses deep learning techniques to produce “human-like” text, and was made widely known when The Guardian published an article written by the AI in September 2020. There has since been talk about the increase in the “apparent applicability” of GPT-3, but this increase is due almost entirely to more data being fed into its system. In short, GPT-3 seems increasingly applicable because it is constantly being trained to “know” and process more information and language; the current version has 100 times more data than the previous one. The AI’s advancement is not due to any unique or innate capability, but to the amount of data it is being shown, just as in the animal categorization example above.
GPT-3 is still a narrow application of AI. Yann LeCun describes it more as a version of auto-complete than as a general AI: “GPT-3 doesn’t have any knowledge of how the world actually works. It only appears to have some level of background knowledge, to the extent that this knowledge is present in the statistics of text,” he points out. It still needs to be taught by humans. This lack of generalized knowledge is what separates it from the general AI we imagine: GPT-3 could probably not have a human-like conversation in real time, and it certainly can’t perform tasks outside of language processing.
Current advances in general AI are being driven by computational infrastructure, for example the amount of data available to GPT-3, not by actual advances towards a human-like AI. This is an interesting step forward, but it doesn’t actually move us towards the goal of achieving general AI.
The path towards general AI
General AI is still an aspiration, and no one is entirely sure when or how we might reach the goal of creating AI that can operate in a more human way. But we want to propose a different approach to the creation of general AI. The idea that human and machine intelligence are the same, or even similar, is fundamentally flawed, and yet developers continue to approach AI as though simply expanding it will result in a closer approximation of human intelligence. This is why creating bigger and better versions of GPT-3 will never result in a general AI.
Human beings excel at thinking in abstract terms. They can be creative, theoretical, and innovative, and in fact these abstract methods are the reason that humans are able to make things, from musical scores and abstract paintings to self-driving cars and skyscrapers. AI, on the other hand, is best at data processing and, so far, incapable of abstract thought. For an AI to complete a task, you must break that task down into pieces of data, because no matter how many layers are added to an AI, and no matter how much it develops its own behavior within its programmed parameters (for instance, making decisions between correct images or numbers), it can only learn based on hard data. It cannot extrapolate beyond the data it has already encountered the way that humans can.
The best application of AI is not in replacing humans at jobs that we do, but in “unleashing the full power of human intelligence” by reducing our workload. If AI could handle the majority of our mundane, repetitive, and even dangerous tasks, we might be free to reach our full creative potential.
To get closer to the creation of general AI, we believe that developers need to keep these goals and differences always in mind. AI may be able to achieve the level of intelligence that humans have, but it will simply never achieve the same type of intelligence. Algorithms are not designed to think and reason or identify hypotheses and causes. Our trained models use statistics to identify specific correlations and then find specific patterns: they cannot operate in the abstract.
This is why we propose a modular approach to general AI. Rather than attempting to make a single model generalize around complex problems — which assumes that AI can “think” the way humans can — we should attempt to combine many specific models that solve complex problems, so that a single AI could take on many tasks. This approach would take advantage of the way that AI works: it can easily be layered to address more complicated problems, as long as the individual layers are carefully built. Although general AI is still only a theory and a human aspiration, this method could get us closer to our goal.