How Smart Can Machines Get
Last Updated on December 14, 2020 by Editorial Team
Author(s): Rakshit Dwivedi
How Smart Can Machines Get
The answer will surprise you
AI is exponentially growing, making it smarter and smarter each day. The prediction algorithm deployed by Facebook to track you and serve you with suitable ads (Ad optimization) works at a staggering 98% accuracy. In other words, you can say that this AI software knows you better than your own family members! We have DeepFace(again by Facebook), which identifies human faces with 97.25% accuracy without factoring in lighting conditions and angle.
What do we get out of AI?
AI has been a pioneer for identification, space research, transportation, and whatnot. Fundamentally, we are already surrounded by AI. Many of us think of artificial intelligence as a giant robot possessing all the human capabilities except it’s smarter. Actually, being smart is true, but AI doesn’t necessarily mean a huge talking robot. AI exists in software as well. Siri, Alexa, Cortana are all technically robots.
The AI boom…
We are witnessing a huge surge in AI models' accuracy. Still, we all live in a deep dilemma of what could be the unprecedented effects of a highly smart and sophisticated robot? Can it really possess the capacity of wiping out human civilization? Is the Terminator going to be true?
Elon Musk explicitly calls AI to be an “existential threat” for us. He goes on to say —
“I think we should be very careful about artificial intelligence. If I were to guess like what our biggest existential threat is, it’s probably that.”
Maybe Elon Musk is right. Maybe he is wrong. No one knows. No one knows because we are not smart enough to predict the future accurately. But we do know for a fact that the current trends in advancements could set up a benchmark to visualize what AI would look like 5 or 10 years from now.
I’ve not known anything to be infinite except the universe. Accordingly, I do argue the fact that a computer can perform calculations under a set of rules of traditional physics. Simply, it means that a computer cannot be infinitely smart because it is limited by the laws of physics.
How smart can machines really get?
We all are guided by the laws of physics. Every physical object that we see around us is guided by the laws of physics, and consequently, so are computers.
A physical computer can carry a fixed amount of calculations because it possesses a fixed amount of energy it could work with. In essence, no object in the entire universe can possess an infinite amount of energy in itself. Everything will run out of energy someday or the other. Like our sun itself will get converted into a red giant in a couple of billion years. Well, let’s not worry about our sun right now.
Now since energy is limited, speed is limited. You can’t feed an athlete 3000 calories/day and expect him to run for the rest of his life without stopping. This is how computers function as well. Let me show you how!
Mathematical equations coming up:
Suppose we have a computer that weighs 1 kilogram. The energy that the computer should possess would be: E = mc², which means m=1kg, and c is the speed of light(=3 x 10⁸ m/s). Putting the values into the equation, we get a number approximately around 8.9874 × 10¹⁶ joules.
Seth Loyd, an MIT professor, laid out an interesting equation that defines the maximum number of operations a computer can carry out given some constraints. The equation looks something like this():
Sorry for making this complicated. But I’ll try to make this as transparent as possible.
Here, what we should be really looking at is 2E/πh. E stands for the energy we just calculated, π is our traditional 3.141592, and h is Planck’s constant, which is 6.62607004 × 10^-34 m² kg/s. Putting those all in, we get a number of approximately 5.4258 × 10⁵⁰ operations per second. Thus, we can define this number as the final or ultimate bound of operations a machine can process at maximum energy. One thing to note here is that the energy value E assumes that it is 100% used by the logic gates to perform operations. In reality, we can never achieve a fully efficient computer system because some energy will always be dissipated in the environment in the form of heat or sound.
By the way, an average human brain can calculate only 1000 basic operations per second. Although a human brain might be slower than an average computer, we still possess the power to build extraordinary things. The same 1000 operations per second brain have the power to build a machine to operate 10 million times faster than us. Who knows if computers can possess the ability to reproduce an even faster tool that works at numbers we can’t even wrap our heads around.
Future of Artificial intelligence
We can never be sure what AI might be like in the future. There’s no right and wrong answer to this, and neither are there any metrics to predict how smart AI can get. But we can know this for sure that AI, however sophisticated, can never be infinitely smart. It will always be capped under the limitations of physics. Nevertheless, while I write this, we never know what physics might have ahead for us. Physics is still being discovered, and unlike what Lord Kelvin once said,
“There is nothing new to be discovered in physics now. All that remains is more and more precise measurement.”
Quantum mechanics opened doors to a new physical world. Everything that we thought was magic was later known to be dictated by quantum mechanics. I hope for a day where AI would be only used to make human lives better and not make them non-existent. In the end, it really depends on how we let AI control us. If we voluntarily provide control of ourselves to machines, we should be expecting a doomsday soon. If we are in control of AI, we can solve some of the great unanswerable questions of the universe. Either way, it will always be pinned upon humans to be the root cause for the actions of an artificially smart robot. So, the takeaway from this would be: it is completely dependent on us how we let AI control us.
 https://arxiv.org/pdf/quant-ph/9908043.pdf — Ultimate physical limits to computation
Published via Towards AI