What is general-purpose AI?

AI is sometimes classified and defined into two types: “general-purpose AI” and “specialized AI”. First of all, AI that we use in real life, for example, AI that automatically learns and processes limited tasks such as quality control in the manufacturing industry and fashion proposals on EC sites, is called “specialization.” We call it type AI. And the other is called "general-purpose AI" that can be seen in science fiction movies. "General-purpose AI" is an AI that has human-like thinking and reasoning functions, and is a humanoid robot that overtakes humans in science fiction movies. He looks and speaks like a human, and is sometimes mistaken for a human in movies.

However, there is a reason why this "general-purpose AI" is not seen in everyday life. For experts, the creation of such AI is considered extremely difficult given the current approach to AI development. For example, Shanghai's "Anbot", a riot countermeasures robot, "Josie Pepper", a humanoid robot that serves as a guide at Munich Airport, Seoul's information robot "Troika", etc. Think about it. These robots have one thing in common. It is programmed to perform a specific function. If you ask "Josie Pepper" a question unrelated to flight or baggage claim, it can't answer because these robots are not "general purpose AI".

These robots do not have the intelligence of humans or the ability to "think" about things that they are not programmed to do. For AI experts, achieving this with current capabilities is “like trying to fly to the moon.” It is possible to make a computer acquire precision technology. Computers can now combine multiple skills to perform more complex tasks. This certainly means more machine intelligence, but again, it's not "general purpose AI". This time, while exploring the reasons, let's explore how we can finally approach "general-purpose AI".

How AI is being developed now

The main reason why “general-purpose AI” has not yet been developed is that AI that we can generally consciously and very precisely perform specific purposes, such as the aforementioned fashion proposal and manufacturing quality evaluation. Because we have been working with a focus on creating . The narrower the model you build, the easier it is to learn and the more likely you are to achieve high accuracy. This is because many models learn by being shown hundreds or thousands of displayed examples and being taught to distinguish or categorize them. For example, an algorithm trained to identify types of animals in images is presented with thousands of images of animals and gradually learns to recognize them. Training a model to distinguish between dozens of animals is more difficult and time consuming than simply distinguishing between cat and dog images. The former requires a much larger number of samples and thus a considerable amount of time. In AI, specificity tends to mean speed and accuracy.

Building a "human-like" model that can perform a variety of tasks is naturally efficient, and has been tried many times before without success.

GPT-3 is an example of an attempt at "general purpose AI" and certainly represents an advance in machine learning. GPT-3 is an example of "conversational AI". It uses deep learning techniques to generate 'human-like' text, and became widely known in September 2020 when The Guardian published an article written by AI. Since then, there has been talk of an increase in the 'apparent applicability' of GPT-3, largely due to the increased amount of data entering the system. In short, GPT-3 is constantly being trained to “see” and process more information and language, making it more and more adaptable. The current GTP-3 holds 100 times more data than the previous version. Advances in AI are not due to unique or idiosyncratic capabilities, but because the amount of data being fed into them is increasing, as in the animal taxonomy example above.

GPT-3 is still a narrow application of AI.  Yann LeCun presents it as a version of autocomplete rather than as a "general purpose AI". "GPT-3 has no knowledge of how the world actually works, and only appears to have some background knowledge if that knowledge is present in the analytic text," he said. pointing out. It must be taught by humans. The lack of general knowledge may be the difference from "general-purpose AI".  GPT-3 will not be able to have human-like conversations in real time, nor will it be able to perform tasks other than language processing.

Recent advances in “general-purpose AI” have been brought about by the computational infrastructure, such as the amount of data available in GPT-3, and are not progress to actually achieve “human-like AI.” . It's an interesting step forward, but it doesn't really move towards the goal of achieving "General AI".

The road to “general-purpose AI”

“General-purpose AI” is still an aspiration, and no one knows when or how we will reach the goal of creating an AI that functions like a real human. However, we would like to propose another approach towards the creation of "general-purpose AI".

The idea that human and machine intelligence are the same or similar is fundamentally flawed, but developers continue to approach AI as if it were simply an extension of AI to bring it closer to human intelligence. I'm here. So even if we remade GPT-3 into a bigger and better version, we still wouldn't get to "general purpose AI".

Humans are great at thinking in abstract terms. They are creative, theoretical, innovative, and in fact these abstractions enable humans to create things, from musical scores and paintings to autonomous driving cars and skyscrapers. While AI is good at information processing, it does not have the ability to think abstractly. In order for AI to complete a task, humans need to break it down into data. This means that even if we add layers to the AI, or let the AI evolve its own behavior within the programmed parameters (e.g. determining whether an image or a number is correct, etc.), the AI will still be able to Because you can only learn by doing. AI cannot infer things that are not written in the data.

The best use of AI is not to do things that humans do for us, but to reduce our workload and "bring the best out of human intelligence." If AI could handle most of our mundane, repetitive, and even dangerous tasks, we might be able to reach our full creative potential.

In order to realize the creation of "general-purpose AI", developers will always need to keep these goals and differences in mind. AI may attain human-level intelligence, but this is not simply the same type of intelligence. Algorithms are not designed to think, reason, or identify hypotheses or causes. Trained models use statistics to identify specific correlations and find distinctive patterns, but they are not capable of abstract manipulation.

Therefore, CrowdANALYTIX/ Macnica proposes a modular approach to creating "general-purpose AI." Assuming that AI can "think" like humans, we should combine many specific models that solve complex problems, rather than trying to generalize complex problems with a single model. , the combined AI will be able to perform multiple tasks. This approach takes advantage of the way AI works: If you build each layer carefully, you can easily layer it to address more complex problems. "General-purpose AI" is still a theoretical concept and is just a human desire, but I am sure that this method will allow us to reach our goals.

▼ Business problem-solving AI service that utilizes the data science resources of 25,000 people worldwide
Click here for details

Source:https://www.crowdanalytix.com/is-general-ai-achievable/
CrowdANALYTIX Resource Library