Let’s imagine that a mystery opponent challenges you to a five-mile race. You don’t know whether they’re a good runner, but you decide you’ve got a 50:50 chance of beating them.
But then you find out a crucial fact: when Mystery Man starts, he’ll be 15 miles behind you. You’d be confident of winning, right? Even I, as an occasional jogger, would be confident of beating a world-class athlete with a 15 mile head start.
But there’s a catch. You’ve assumed he’s running – but what if he was actually going to be riding a motorbike. Are you as confident now?
It takes me about 45 minutes to run five miles. Even a modestly-priced motorbike can do that in well under three minutes. It could start over 70 miles behind me – and still win comfortably.
And so it is with artificial intelligence.
We look around us, and we see machines that are dumb. They can add and subtract, display pictures and – at a push – recognise some of what we’re saying. It’s just not that impressive. And because of this, we don’t get a deep, emotional sense of the potential benefits of AI – or the threat.
Let’s look at the numbers, to see how close we are to the motorbike example above.
We can start by trying to assess the capacity of the human brain to improve. It’s been 65 million years since the extinction of the dinosaurs. At that time, the ancestor of humans had a level of intelligence that would perhaps be roughly comparable to a modern rodent. A mouse has four million cortical neurons, while a human has 23 billion. Crudely, that means we’re 5,750 times more intelligent. To put it another way, you’d have to double the intelligence of a mouse around 12 times to get to the intelligence of a human. Assuming that each such doubling took around the same amount of time, it took evolution about five million years to double the intelligence of a mammal. Therefore, on an investment timescale, human intelligence is essentially static. You’re not so much running a race against AI, as sitting by the roadside having a picnic.
Of course, there’s a chance that genetic engineering will make a difference. But we’re looking at a range likely to be within the norm of human variability. This is because we don’t have a super-intelligent sister species from which to poach genes. So, over coming decades, we could potentially expect humans to use genetic engineering – and get a few grades smarter. But it won’t turn the world upside down. Even the most intelligent people today have an IQ less than double that of the average person. So genetic engineering might give us a small push in our race, but it’s questionable whether it would get us much further – without fundamentally changing us as people. Humans with watermelon-sized brains are unlikely in my lifetime.
Meanwhile, AI is coming up behind us – and it’s coming very, very fast. You’re probably familiar with Moore’s law. In its purest form, it states that the number of transistors on a chip doubles every 18 months. It’s an exponential curve – just the kind of thing we look for at Exponential Investor.
Moore’s law has to break eventually, because there’s a real, physical limit to how many transistors can be forced onto a chip. But this doesn’t mean it’s wrong, in a more general sense. We can expect that Moore’s law will hold, in some form, for a long time – giving us electronic “brains” that roughly double in capability over an 18 month cycle.
Now I wouldn’t like to speculate on what the precise doubling rate for artificial intelligence capability is. But I see no reason to assume that we’re going to see a fundamental disconnect between Moore’s law and the emerging power of AI. It might not be 18 months, but it won’t be 100 years – or even ten.
And that’s a big deal.
Compare this to our example above: five million years to double human intelligence. That AI motorbike is coming at us very, very fast as we run. In fact, a motorbike isn’t an apt analogy at all. AI is approaching us a rate around three million times faster than we’re running. The fastest man-made craft is the Juno space probe, which travels about 15,000 times faster than I can run. By contrast, the rate at which AI is advancing on human intelligence is 200 times faster even than this. It’s an incredible, staggering rate of advance. Once you put it that way, it become obvious that AI will eventually catch up – no matter how fast humans “run” in this race.
But surely there’s nothing to worry about for a long time? Wrong.
Today, we’re at the stage where we can exactly model the brain of a very simple worm. Let’s assume that today’s accurately-modelled worm is the cutting-edge of AI – and note, worms can’t drive cars, but computers can. We don’t need to exactly model an animal brain, in order to build an intelligence that outstrips it.
But we’ll keep things conservative and start with the roundworm’s brain – all 302 neurons of it.
Double the intelligence of this artificial worm once every 18 months and in around 40 years, you’ll end up with a machine that’s as intelligent as a human. Yes, it’s only around a single human generation before AI will be as smart as a human is today – based purely on extrapolating long-established trends. That’s the power of the exponential.
And, if you think that is a big deal: you ain’t seen nothing yet. Later this week, we’ll return to the subject – and I promise you: I’ve got something that will blow your mind…
Category: Artificial intelligence