Wanna become a data scientist? Checkout Beyond Machine!
Why We Need General AI (and Why We Are Not There Yet)
Most people talk about AI in very broad terms. The reality, however, is the artificial intelligence we have now is very different to the AI we see in the movies, i.e. AI where robots and machines are conscious, can deal with a myriad of situations, and have emotion.
Specifically, the AI we have today is known as Narrow AI or Weak AI. The AI we want to get to in the future is General AI, also known as Strong AI. What’s the difference between the two, and what is needed to get to Strong AI?
The Difference Between Narrow AI and General AI
Before delving further into General AI, it’s important to understand how it differs from Narrow AI.
AI tools like Siri, Google Assistant, and Alexa all use artificial intelligence and machine learning. However, each one of these platforms focuses on narrow, specific tasks.
When you ask it something it understands, it’s great. Ask Siri what the weather is like, for example, and it will know how to answer you because it has been programmed to perform this task.
When you ask Siri something more challenging, philosophical, or complex, however, it has no idea. Sure, it might turn your words into a Google search, presenting you with the results. It won’t know the answer to the question, though. In fact, it doesn’t even understand the question.
General AI is very different. A machine with General AI will be able to do almost anything by using similar skills and traits found in humans. This includes using knowledge of topics in one area to make decisions on something completely different. It will also be able to reason, plan, show innovation, and demonstrate problem-solving capabilities.
In other words, General AI will be able to handle a new situation it has never experienced before.
For example, a Narrow AI machine might be able to book a restaurant reservation because that capability is built into its dataset. It uses AI to learn your voice and how you speak, it translates those words, and makes the reservation.
What happens when you ask that Narrow AI machine to make an appointment at the dentist for your daughter? A human understands the mechanics of doing this are very similar to booking a restaurant reservation. Unless it has been trained to do this specific task, a narrow AI machine will have no idea what to do.
A General AI machine, on the hand, will understand the similarities between the two requests and it will use its learnings from making the restaurant reservation to make the dentist’s appointment. Crucially, it will know what parts of the restaurant reservation process are not relevant to booking an appointment at the dentist. How it could do something like that? There is already a hint behind those capabilities in algorithms such as word2vec.
Word2vec turns words into thought vectors, and then these vectors can be combined or manipulated in an intuitive way, similar to how us humans approach language.
Is Narrow AI Important and Why Is It Important to Get to General AI?
While many Narrow AI solutions are very advanced by today’s standards, we still have quite a way to go to get to General AI. That said, Narrow AI still offers considerable benefits. We just need to take a look at some of the machine learning applications around us to get a taste.
For a start, it saves us time plus it’s better at doing certain tasks – data-driven tasks – than humans. For example, autonomous cars have the potential to be safer (because they don’t get tired, they don’t become distracted, and they don’t drink and drive), plus traffic jams are likely to be much less frequent when all the cars on the road are autonomously driven.
Narrow AI is also better at analysing data, improving productivity in business, and making business processes more efficient. It also has applications in healthcare, manufacturing, and product development.
General AI will take these benefits to a whole new level, reducing the number of machines and platforms we need while also performing tasks not only better than humans, but also better than Narrow AI machines.
This will make businesses more productive and efficient, improving profits. Health outcomes should improve, society should become more efficient, and we should have more leisure time. There are challenges, however. For example, humans are already losing jobs to machines who can perform the same tasks better and cheaper. This trend will only increase in the future.
Apart from the human impact of these changes, many are concerned with things like public revenues, i.e. if people don’t have jobs because AI machines are doing all the work, who is going to pay taxes? Some say the robots should pay (in a manner of speaking). Others say people will still work, just not in the jobs they do today.
However it shakes out, the future will look very different because of advances in AI technology.
Why general artificial intelligence is not there Yet?
There are several reasons why General AI is not yet a reality. However, there are various theories as to what why:
- The required processing power doesn’t exist yet. As soon as we have more powerful machines (or quantum computing), our current algorithms will help us create a General AI.
- Programmers have to figure out how to get machines to learn as humans learn rather than simply working from a predefined (and, by definition, narrow) dataset. This requires a fundamentally new way of thinking about machine learning (e.g. Friston’s free energy principle)
- Maybe we just need better algorithms. Judea Pearl for example has argued that until machines understand causality, we won’t have real AI.
Judea Pearl is one of the most prominent computer scientists of our time.
As a result, we’re probably many decades away from General AI, despite the advances made in AI to-date and expert opinion is divided as to how close we are. Broadly speaking, we’re on the right track, though. So, it remains to be seen what the next evolution of intelligent systems will look like. I personally believe that some of the work that is key in building general AI, is reinforcement learning, and algorithms such as deep q-learning that Deep Mind has investigated.