The Deadly Gamble on Super A.I.
Rob Reid

Put me in the category of “people who would be overjoyed if any of our biggest problems were strong AI”

This is your brain on AI

I found this article charming with a lot of good ideas — but worrying about rogue AIs in 2018 is like worrying about asteroid strikes when you are a two-pack-a-day smoker.

I’d be overjoyed if I thought that human civilization and technology would continue progressing until we had actually intelligent machines. Surely beings smarter than we are could hardly be worse stewards of our world than we ourselves have been?

I feel strong AI — “machines that can reason and think like humans” — is possible, but that it will take a very long time — and unfortunately, our civilization does not have a very long time — because the progress towards strong AI has been very slow. Indeed, I first took an AI course in the early 1980s, and none of the problems that AI was interested in at the time have advanced significantly — particularly in comparison to, well, all the rest of the programming world.

We all know there has been immense progress in fields like machine learning, but this is is nothing at all like strong AI, and more, gives us not the slightest hint or clue as to how to proceed towards strong AI.

Consider “The story problem” — you give a computer program a simple story, like, “Mike goes into a restaurant and orders a steak. It is burnt. He yells at the waiter and leaves”, and them you ask a question about the story — like “Did he pay?”, something which isn’t explained specifically in the story, but where any person would say, “Probably not”.

Well, the state of the art in 1982 were programs that could do a mediocre job on that task in one domain — like “restaurant stories”. The programs were smart enough to understand that the “it” that was burnt is the stake and not the restaurant”, that “he” is “Mike” and not the waiter — but anything more complex is lost, and in order to answer questions like “Was Mike happy?”, “Would he recommend the restaurant to his friends?” you’d essentially have to hand-program all these rules. And there are just too many such rules…

And there have been little progress on this problem since. And this is the sort of basic thing that any child can do. You would never call a computer that couldn’t understand elementary school-level stories and answer questions about them “intelligent”.

Again, I don’t think strong AI is impossible — it’s just hard, and there needs to be a whole new line of attack, and I can’t even speculate on how that would go. Perhaps there’s some subtle trick someone will come up with tomorrow, and voilà — perhaps it would take decades, or centuries. Progress to date hasn’t been encouraging, and there’s nothing that lots of CPU power will do to help. I remain hopeful, but hope doesn’t get the crops harvested.

Unfortunately, we don’t have centuries. A much bigger problem is facing us — the destruction of our ecosystem. Between climate change and the ongoing sixth mass extinction we face real, immense problems from the degradation of our own environment, problems that already kill millions of people a year and will only get much worse as feedback effects manifest — and our leaders seem uninterested or actively hostile towards action of the appropriate magnitude.

So if we get past this bottleneck to strong AI, I’ll be elated.

I do declare that sometimes I wonder if people like to fantasize about these distant, fantastical threats so they can dismiss them sleep at night and not worry about real, boring, grimy threats that look like they’re really happening.

And you can relax also about the evil AI or the Basilisk eating the world— if strong AI does happen, it isn’t just going to pop accidentally, magically from someone’s Bayesian spam filter neural net getting it on with a support vector machine learning model. It’s going to take a brand-new approach entirely and likely many decades of sustained work, and “moral values” will definitely be part of that gargantuan effort, if only because of all these works of fiction — after all, it’s been almost two hundred years since Frankenstein.

Let us hope humanity sees that day!