Terrifyingly quick learners

Humans are (currently) considered smarter than computers. We’re better at ‘one-shot learning’: guessing the right action in previously unseen scenarios. We can do it with next to no training data. Toddlers only need to touch a hot stove once to learn not just never to touch one again, but how to recognize them even if they might be different shapes, sizes, colors, and locations. Computers are better at computation: they can try many million more combinations in an instant, than any one human can in a lifetime. That’s just one computer, but they’re better at handing off tasks too: computation can be split across multiple computers to be worked out in parrallel, with far more reliability than the mess we get trying to coordinate multiple humans. The way computers get close to human levels of ability in many domains is through sheer brute force: more training data, bigger models, more cloud computing power.

The human brain is the only known general intelligence in the universe (so far). Our brain improves on evolutionary timelines – over thousands of years – so we’re still operating on pre-historic hardware. We rely on our software – shared culture, knowledge and technology – to keep us functional in the modern world. To take a list from the famous Robert Heinlein quote, human beings are able to “change a diaper, plan an invasion, butcher a hog, conn a ship, design a building, write a sonnet, balance accounts, build a wall, set a bone, comfort the dying, take orders, give orders, cooperate, act alone, solve equations, analyze a new problem, pitch manure…” and of course, “program a computer”. It’s worth remembering that just because we invented the computer, that doesn’t mean we’ll forever be superior.

Computers improve exponentially; Moore’s Law is a historically-reliable rule describing how every year we get 60% more computation than the year before. We’re nowhere near human-level — 10 quadrillion calculations per second. Computers are at about 1,000/th of that: as smart as a mouse. This doesn’t sound like much, until you know we were at a trillionth of that in 1985, a billionth in 1995, and a millionth in 2005. Humans always underestimate when estimating exponential growth, but imagine you’re filling a glass of water, exponentially. It’s an indescribably small amount of water at first, and you slowly increase the amount you’re pouring over the course of 60 minutes. At what point is the beaker half full? The answer is counterintuitive and surprising: 59 minutes! Up until the stroke of midnight we’ll be able to claim superiority, after which, all hope of keeping up will be almost instantly lost.

We won’t have time to prepare for computers to surpass our intelligence, they’ll fly past us, speeding up as they go by. Already the number of papers on AI are growing exponentially, and researchers are using AI to summarize the new AI papers that are coming out, as well as predict and suggest new areas of research. AI code assistants are reportedly writing up to 40% of the code produced by software developers using it, with a realistic target of 80% of code being written by AI in 5 years. There have even been successful attempts at using AI to program better AIs, making fundamental discoveries of faster ways to do the linear algebra required to train neural networks. This is the beginning of the end for the human brain’s superiority complex.

Computers will quickly become so intelligent that we likely won’t be able to comprehend their intelligence. How do you comprehend the goals, incentives, culture, of something twice as smart as you? 100x as smart? We’d be like ants trying to guess what humans might be thinking about. If AI does remain benign, we’ll have unprecedented power to solve the world’s problems. If not, then I, for one, welcome our new robot overlords. If there is a malevolent AI in our future, it might go easy on those who helped bring it into existence. Our genes might not live on past the singularity – the point when computers become powerful enough to simulate a human brain – but our memes certaintly will. Should we build a brain-computer interface, we can come along for the ride. The question is not if we can beat them, but can we join them?

About the Book 📕

← Back to Essays 📝