When will AI-powered machines take over humans?

We are surrounded by machines that can make decision, sense environment around us and adjust its behavior without human intervention. At the heart of these digital smart devices are Artificial Intelligence and Machine Learning - the ability of a machine to perform cognitive functions we associate with human minds, such as perception, reasoning, learning, judgment, creativity and problem solving. With all the optimism artificial intelligence brings; what come in parallel are the concerns that it will create mass scale unemployment, end of humanity, or make human slave of machines. No matter which side of argument you are on AI’s impact on world; it is relatively easy to realize that the tendency of human to seek technical solution for a real-world problem will keep pushing the capability of AI in general. 

The evolution of Intelligent machines is not very different from any other phase of technical evolutions in past where capability of machines continuously improved, initially intended for increased productivity and eventually resulting in full industrial revolutions. However, one very interesting dimension of evolution of AI is that – in past technical evolution was always instrumented by human tenacity to find solutions using machine. Whereas AI-powered machine can theoretically evolve itself without help of Human. For example - a steam engine’s throughput was improved after R&D done by humans to make efficient design changes. Pentium 1 microchip cannot do research and development to design a faster Pentium 2 microchip. But an AI-powered machine can keep learning and improving by itself over time. An AI designed to negotiate online bids can just keep learning and become more efficient without any help needed from human (at least in theory). Or invent their own language which Human doesn’t understand. Scary thoughts .. right? 

There are many unknowns on what and how AI will evolve over time, however AI will evolve broadly in three distinct phases. Each type of AI will have its own evolution path and timeline. It will be dependent of the choices made in previous phase. 

Artificial Narrow Intelligence (ANI)– Narrow AI is a machine designed to do one task very well. These machines are not programmed to do any other work no matter how easy or difficult it is. For example – an AI system designed to detect a potential fraud in claims processing cannot consume image of simple traffic light and say whether its green or red. Same way a sophisticated face recognition AI system cannot detect a fraud in medical claims. Even self-driving car is a just a combination of numerous narrow AI, each AI trying to do one specific task better than human.

Narrow AI has already arrived and being refined at blazing speed every passing day. Machines have already started replacing humans for repetitive tasks and the sophistication of task is just going to increase.

Artificial General Intelligence (AGI)

When you think of General AI (AGI) is when you start appreciating the beauty of Human brain. AGI also considered as “True AI” is when a machine can perform an intelligent task like any normal human. General AI is the dream of AI enthusiast and is a stopping point for all AI realists. General AI has always been illusive, from many years futurists have been practicing that General AI is around the corner. However deeper you look, the more you realize how difficult it is for a machine to achieve AGI if ever possible. In 2007, Apple co-founder Steve Wozniak prescribed “coffee test” to validate if a machine has achieved General AI. Coffee test is walking into a new American house for the first time and making coffee – which means locating kitchen, finding coffee machine, finding coffee powder and water, finding a mug, brewing coffee by pressing right button of the coffee machine of unknown brand.  Steve Wozniak predicted that these skills cannot be programmed and has to be learned which only living brain is capable of. For a machine to be able to pass “coffee test” is extremely difficult which most human can do with much ease (Quality of coffee is not a parameter here, otherwise most human including yours truly will fail J). If you are thinking of a machine specifically designed to pass “coffee test” you are still thinking of a combination of ANI not AGI.

A human brain can juggle between so many unrelated task, utilize social skills and wisdom which a algorithm will have very difficult to replicate. We may be decades away from AGI, when AGI is achieved that will certainly have massive effect on humankind, jobs, social structure and virtually everything in world as we see today.

Artificial Super Intelligence (ASI)

Super AI is when a hypothetical machine can do anything and everything better than human being. Stephen Hawking argued that “humanity will end”the day when Super AI is developed. Similar is the warning of Elon Musk when he says “AI is more dangerous than nukes”. Dark predictions hold true if you believe that a true super Artificial Intelligence will be able to create a machine with a complete replica of human intelligence. The problem is when we consider word “intelligence”, our mind subconsciously associates it with “positive” connotation. However, in true sense, a super AI will have all positive as well as negative intelligence of a human. So a true super AI machine will have creativity, analytical genius, emotional intelligence and social dexterity. At the same time super AI machine will have feelings such as fear, anxiety, jealously, revenge, racism, and bias.

We are already living in a world full of Narrow AI. Machines are going to be able to perform more complex task every passing day. By combining multiple narrow AI in one giant AI system, complex ANIs such as self-driving car, clinical support system, robotic surgery etc. should be a mainstream reality soon. Transition from narrow AI phase to general AI phase is tricky and might take decades, if ever true General AI becomes possible. This transition is also path dependent - meaning decisions made for ANI will influence when, what and how general AI is developed. Transition from General AI phase to Super AI phase should be fast, may be few years, few months, or just few days. In fact, general AI to super AI transition could be as fast as snap of a finger and might not need any human intervention at all. A general AI machine might just outlearn itself into super AI. A robot which will becomes too efficient in negotiating and writing screenplay, will also learn plagiarizing and taking credit of someone else writing. A machine which can operate everything in your house without you training will also learn and find a reason to trick you out of house and not let you in.

Artificial super intelligence opens the Pandora box. If ever a Super AI machine is developed, how do you stop that machine? An ASI machine will theoretically be able to think before your brain that you are about to shut the machine and hence activate the defense mechanism.  Human intelligence is roughly between 80-150 IQ level and is limited by neuroscience. If matching human intelligence is what Artificial Intelligence is planned for and a Super AI machine with 150 IQ is developed, how do stop it at 150? Why machine cannot develop 300, 3000 or 30,000 IQ level, if all it needs is few more server in cloud to add the processing power. This is where path dependency comes into picture, a decision made early in evolution of AI will decide what shape Super AI will take. One of the most sane voice in field of AI and Baidu’s deep-learning genius , Andrew Ng summarizes this future very well “worrying about General or Super AI is like worrying about over population on Mars before even setting foot on it”.

I welcome your views, feedback, agreement and, of course, disagreement.

Disclaimer - Views are personal.