Hacker News Reader

Comment by nostrademons | Hacker News Reader

A number of us live in Silicon Valley or work daily with AI. These aren't science fiction technologies for us: they're, real, tangible cars that you see daily on the roads, or software libraries that you can download and use yourself. I can understand why they'd seem like fantasy if you live in some far-off place, but when you're stuck behind a self-driving car going 25 mph on El Camino, there's pretty concrete evidence that self-driving cars are here.

I do think that AI and particularly deep learning are overhyped. It is a big deal, but it's not as big a deal as many people make it out to be. It's much like the Internet in 1998; it works, you can use it now, and you know that it's going to drastically increase the reach of computing power, but people are talking like that's all going to happen in 1-2 years when really it'll be 2-3 decades.

View on HN

I guess my concern isn't that these are "science fiction technologies" but that they're rather simple. AI has a history of over-promising and under-delivering and even if a few "self-driving" cars have made it on the road, how long until people begin trying to use them in scenarios they're not really ready for? All of that progress can disappear overnight if there are a few bad accidents.

I guess when I look around, I don't see a comparison between AI now and the Internet then, I see a comparison between AI now and AI then (70's).

AI has this effect where whenever an approach starts working, it ceases to be AI:

https://en.wikipedia.org/wiki/AI_effect

It's worth remembering that many of the actual technologies developed in 1970s AI - A*, beam search, pattern matching, constraint solving, rules engines - are now basic CS algorithms that most programmers are just expected to know, and they have continued applicability to a wide variety of domains. We just don't call them "AI" anymore, because the technology has been commoditized enough that they are "business rules", or "single-player games", or "layout engines", or other examples of the actual problem domain that's solved.

Ditto 1990s AI - fuzzy logic, SVMs, Bayesian networks, collaborative filtering, perceptrons. It just got good enough that it became the standard way you solve a certain class of problems with computers, and now it's called "fraud detection", "Zestimates (tm)", "recommendation engines", "stream ranking", and so on.

Deep learning is the same. The hype cycle will eventually bust, because all hype cycles eventually bust. But in the meantime, it has opened up whole classes of problems in computer vision, speech recognition, recommendation, machine translation, and several other domains that were not amenable to computerized treatment before.

I guess that's sort of my point though isn't it? A, SVMs, etc. aren't really called AI because they're not. Because AI carries a connotation and once you understand A it's obvious that it's not really AI. Won't the same be true of today's neural networks? Granted they've been given a name which should add some longevity but when a layperson thinks "AI" they think a robot which acts like a human, they think AGI. If someone had an AI algorithm which could start to demonstrate more AGI leanings I bet the AI effect wouldn't apply.

When cars came out they killed quite a number of people in accidents and mechanical failures. They were just so much better than horses we kept on improving them.

That's how I see AI. People will surely get injures and even die in Accidents. We are a humanist society where human life is valued above everything else, so I'm very optimistic that AI will get better and give their humans more divine powers and longer life over time.

Traffic in the US also seems vastly easier than in other parts of the world, mostly because roads are wider. When i see the some of the chaos in cities like Paris, Barcelona, Rome and the narrow roads in many places, i can only come to the conclusion that it will take a long time before an AI will be able to handle that. Highway/intercity traffic sure, that should be solved in a few years, but the rest will take some time. Even in Germany we have lots of backroads on the countryside that are so narrow (and have no markings), that you have to go off the street with half of the car if someone is coming from the other direction.

Even assuming your conservative adoption/relevance curve of the technology itself, I don't think we need too many more incremental steps before society has to stand up and take notice of an unemployment problem that can't keep up. AI / automation might not turn our day to day lives into the Jetsons in 5 years, but if it gives us 25%+ real unemployment we're still going to be in for a ride.

Reply on news.ycombinator.com