The general truth that “anything a human being can do, AI can do better” is in the water supply, and of course that isn’t quite true.
When it comes to narrow, focused tasks, there are many, many things AI can already do better. There are also many, many things it can’t.
And the part of this truth that needs to be brought into the light is not so much that machines surpass humans in a few areas – bulldozers have been outlifting Olympic weight-lifters for quite some time now – but just how much better it’s possible for them to get.
And then there are the things an AI can do, cognition-wise, that a human being will never be able to do.
Here’s a list of ways in which AIs can already exceed human cognitive performance.
The number of neurons
A deep learning model contains an artificial neural network – a learning mechanism that possesses a vast number of nodes, simulated neurons, each of which is connected to many others (simulating the synaptic connections of actual neurons). These neurons, heavily layered, are the hard physical limit on the capacity of the network.
The human brain contains roughly 90 billion neurons. Each of those neurons can be connected to 1,000-10,000 neurons, meaning that there are orders of magnitude more connections than neurons – somewhere north of 10 quadrillion.
A digital neural network has no such limitations. On today’s computers, neural models with more than 100 trillion neurons – with a correspondingly astronomical number of connections – are already becoming common.
It is now possible to build an artificial neural network, for the purpose of creating a deep learning AI model, with 100,000 times the complexity of a human brain.
What’s the consequence of that difference?
Superior pattern sensitivity
One of the core characteristics of intelligence is pattern recognition. All creatures – mammals in particular – are able to navigate the world and survive because they can detect patterns in what they observe in the world, via the senses. Day and night... the presence of food in some places but not others... the motions and behaviors of other creatures. Our own pattern sensitivity is exquisite, across our senses; that the senses themselves are not as acute as those of some animals does not diminish the importance of that pattern-sensing ability.
Because AI neural networks have far more neurons, they are able to detect patterns in their input (usually large bodies of data) with far more sensitivity and depth than a human being. The popular example is in radiology: AIs can read X-rays far more accurately than even the best, most experienced human radiologist. This has been true for years, and will only get better over time.
One reason for this superior performance is that the AI has practiced radiology for far longer than any human.
20,000 years of experience
Above, it is mentioned that when the engineers at DeepMind trained AlphaGo Zero, they gave it the equivalent of 20,000 years of human experience playing the game – and did so in just a few weeks. Almost any neural net-based deep learning model can be given this depth of experience: many thousands of human lifetimes. How skilled would a human doctor, for instance, become in the course of such an epoch?
An actual human doctor, on the other hand, will invest 12 years, at most, in training – and would be limited to some subset of 24 hours a day, and a subset of seven days a week, in the effort. Training and learning continues, of course, at a lesser pace, but even so, the total is miniscule compared to what an AI can do. Almost every other skill set that exists requires less.
And, of course, AIs are not limited to their own learning.
Literally sharing learning and experience
with others of their kind
Human doctors, of course, consult with other doctors – checking their knowledge, soaking up new knowledge – but deep-learning AIs can already do more. It is now possible to combine networks, so that both absorb each other’s training – and for a neural network, fully trained, to train other networks.
Put another way, deeply-trained AIs can transfer their training to other AIs. This technology is still young, but its direction is toward an AI universe where every AI of a certain type shares knowledge with all its peers.
Imagine if all the heart surgeons in the world were telepathic, and had access to each other’s experience, knowledge, and skill. Imagine every heart surgeon having all the experience and skill of all heart surgeons.
That’s what happens next.
And that’s not all...
AIs can be endlessly cloned
Let’s say our heart surgeon, who not only has 20,000 years of experience as a heart surgeon, and who shares the knowledge and experience and skill of all heart surgeons, can be cloned on demand, in seconds – that he could step into a device like Star Trek’s Transporter, and a copy would pop out.
It goes without saying that this is already trivially simple, and breathtakingly inexpensive, to do with highly-trained AI models, their algorithms, and their instantiated apps.
Now let’s imagine that we’re not talking about a heart surgeon, or any other medical or legal or financial specialist – but a military commander, a battle strategist.
Imagine an AI that was trained as a battle strategist – with a cognitive capacity 10,000 times that of a human military commander, with 20,000 years experience, able to detect patterns in enemy activity with orders of magnitude more sensitivity than a human, who had access to all the experience and knowledge and skill of a million other AI commanders?
Would any living human military commander be willing to face such an AI in combat? Would any human army be willing to follow a human commander into battle against a force commanded by such an AI?
And that’s before we get to collective expertise.
LinkedIn, the biggest social network for professionals, claims that the posts, articles, and other human artifacts on its site represent a collective 10 billion years of human business experience.
Now imagine that a deep learning AI trains on LinkedIn’s collective data. Imagine that it discovers patterns in the data none of its contributors could ever detect. Imagine that this AI is placed in charge of a Fortune 500 company.
Would you want to be the competitor of such an AI?
Think before answering any of these questions. Because they are all real-world questions that will soon demand answers.
Comments