Table of contents
The following image shows what Artificial General Intelligence (AGI) looks like according to a narrow AI.
The image was generated from nothing more than the text prompt "artificial general intelligence" and the text-to-image model Stable Diffusion on Hugging Face.
But what is an AGI? And what is a narrow AI?
In this article, I am going to introduce you to the concepts of AGI vs. narrow AI. I'll talk about my personal motivation to build an AGI, and share some thoughts about the current progress in research and about what might be needed to succeed.
Strong AI versus Narrow AI
The term AGI, which can also be referred to as strong AI or full AI, is often defined as follows:
"AGI is the ability of an intelligent agent to understand or learn any intellectual task that a human being can."
In 2022, we are still far away from creating humanoid robots that have an AGI. But if we could build them, they would be able to:
- have a human-like conversation on any topic
- find out how to make coffee for you at home
- learn about any topic and pass a university test about it
- perform an economically important job at least as well as a human
- understand and solve any task that a narrow AI can solve
In contrast to that, narrow AI is not intended to have general cognitive abilities. Rather, narrow AI is any program that is designed to solve a specific problem.
Examples of tasks which have been (more or less) solved by narrow AI are:
- detect and label objects on an image
- perform sentiment analysis on a text
- play chess at a super-human level
- self-driving cars
- generate AI art, as seen above
There are many tasks at which AI already excels today at a super-human level.
So why would you build an AGI?
There are a couple of reasons why.
Why Build AGI?
If you've seen movie such as The Matrix, 2001: Space Odyssey, Ex Machina, Her or War Games, then you know that AGI is a popular and exciting topic in science fiction.
It's a topic that fascinates and frightens at the same time. AGI could be the best or the worst thing happening to humanity.
Humanity is making progress in many scientific domains. We're working on curing rare diseases, fighting world hunger, and saving our planets from all kinds of threats. We're working on those issues as fast as we can.
What if a super-human intelligence could solve those problems for us?
A system with strong reasoning skills, vast amounts of information, and an enormous computational power, might be able come up with solutions for many of our problems. It's hard to imagine all the possibilities how to use it for good. It might be our last invention.
"solving intelligence, and then using that to solve everything else."
For me personally, artificial intelligence in general has been fascinating to me for decades. It started when I was a kid as I was watching some of the movies mentioned above. What if intelligent robots could become real? Would they improve our lives? Would they take over control and become malicious?
I don't know.
But what I know is that I can hardly imagine anything more exciting that working on finding out.
Forecast
I don't feel qualified to make a forecast about when, if or how AGI is going to arrive. There are many experts who shared their opinion. According to a study, there is "no difference between predictions made by experts and non-experts" and since pretty much the beginning of AI research, "there is a strong bias towards predicting the arrival of human level AI as between 15 and 25 years from the time the prediction was made".
So I'll put in my two cents anyways.
The past decade in machine learning has been incredible. There have been major advances in computer vision, NLP and Deep Learning in general. The AI-generated image at the beginning of this article is just a recent example of what has become possible.
The reasons for this incredible progress are mostly:
- data storage has become cheap
- huge datasets have become available
- computational power has increased exponentially (Moore's Law)
- GPU acceleration for Deep Learning
- a hype around AI/ML, resulting in more investments flowing into research
There are no signs that this trend is going to cease. I'm expecting more amazing results in narrow AI research in the upcoming years.
But how are we going to reach AGI?
One option is that a narrow AI becomes so advanced that AGI just emerges by itself. Recently, I was reading a story saying that a Google engineer claimed that its chatbot has become sentient.
Large language models like GPT-3 or LaMDA, the one mentioned in the story, are rather great pattern detectors than thinking agents. They are unable to reason, plan, learn, play games or perform any task other than generating plausible text.
Maybe, if those models become larger and more complex, at some point AGI abilities will emerge. But I wouldn't bet on that.
There are all kinds of approaches to creating AGI.
I believe that the successful ones won't be the ones that focus on a single large model trained on a giant dataset. Instead, the usage of many methods and scientific fields will be required:
- insights from human psychology and neurology
- combining different machine learning approaches
- supervised learning
- unsupervised learning
- reinforcement learning
- using an evolutionary approach
- creating a hybrid model with both neural networks and the symbolic approach
- taking biological plausibility into account
Conclusion
That's just what I think about AGI.
What do you think?
- What is needed to create AGI?
- When will it happen?
- Will it happen at all?
- Would it be good?