Hacker News new | ask | show | jobs
by blackhole 3263 days ago
I appreciate the point of view this post is providing. It is always important to think about the potential limitations of AI, and the fact that once again, we may hit the ceiling of our current attempts at AI much more quickly than we realize, and no singularity will ever occur. We need to think about how to solve our problems realistically, now, without waiting for a godly super AI to come solve them for us.

However, what's bizarre is that he is painting a world that is already wrong. In particular:

"We won’t have massive, perfectly coordinated networks with optimised flow and distribution — think traffic networks (those self-driving Teslas that act as taxis when you don’t need them)"

But... we will. We're already making it. It might not work very well at first, but automated cars are a real thing, and they are going to happen. We have preliminary automated cars right now. Perhaps he is instead claiming that our autonomous cars won't scale, but this too makes no sense. Of course we can make them scale. Once we've solved the hard problem, which is actually driving the cars around and not hitting things, solving distribution and flow is almost trivial. We have all sorts of systems for solving distribution problems and finding maximal flow along a network. There's even an entire class of algorithms to solve it with[1], which we currently use, right now, to solve things like scheduling airplane flights.

And then he brings up a point that seems completely perpendicular to the entire rest of the post:

"We won’t have total surveillance (à la Reynold’s Mechanism or Brin’s Transparent Society)"

This has nothing to do with AI and everything to do with cryptography. Whether or not the singularity does or doesn't happen is completely irrelevant. Indeed, it currently looks like we're headed towards having total surveillance with or without the singularity, unless we do something about our privacy laws.

While we shouldn't assume AI will fix all our problems, the examples provided in this post are bizarre, to say the least.

[1] https://en.wikipedia.org/wiki/Maximum_flow_problem

7 comments

Your first paragraph is spot on.

The "singularity" is really just 17th-century metaphysics. It's Rene Descartes all over again imbued with religious undertones of salvation and transcendence. Most of the "aspirations" of its occurrence are nonsense. Kurzweil has fashioned himself like an AI prophet. For some reason, it has caught on among people, some very smart, and has lead to this general passive-ness to solving real problems today. The "singularity" becomes the answer for everything, "it's definitely coming!"

The only problems I'm interested in solving today are the ones that will make me rich enough that I don't have to work when the robots take over.
The author made the mistake of citing specific examples that weren't very good. It's odd to me that's there's this conflation of ai with the singularity. Strong ai is only one path to the singularity, others being human augmentation, and group minds. Also, the idea of this all being a 'failed dream' was addressed by Vinge in this talk[1]. He picks a metric, then tracks how it would develop in the case of a singularity never happening. Very enlightening to see his thoughts on this.

[1] http://longnow.org/seminars/02007/feb/15/what-if-the-singula...

My pet theory is that if everything goes as well as it can go, people thousands of years from now will call the period we are currently living through the singularity.

I mean, if you look at the definition of singularity that is implied by the site behind your link -- "self-accelerating technologies will speed up to the point of so profound a transformation that the other side of it is unknowable" -- then to be quite frank, that has basically already happened.

BTW, I'm not trying to say that the singularity has ended already. It's a development that started 1-2 centuries ago and is currently ongoing, and pedants of the future will quibble that it's not truly a singularity (but then again, true "naked" singularities simply don't exist in nature, so that point is moot).

> "self-accelerating technologies will speed up to the point of so profound a transformation that the other side of it is unknowable"

It is not even the first time this happen. There was something fitting this description starting the Neolithic, that people are not sure what exactly it was. There was writing and the big civilizations that happened through it. And there's the scientific/industrial revolution that we can argue endlessly if it is or isn't just a part of what is happening now (but falls just outside of your timeframe).

Re: writing. Lots of the oldest clay tables are rants about how 'our laws are the best' and 'other cities' laws will make you work too hard'. Apparently writing really took off when it got used for PR!
> solving distribution and flow is almost trivial. We have all sorts of systems for solving distribution problems and finding maximal flow along a network. There's even an entire class of algorithms to solve it with[1],

I would not be so confident with that. Yes, in a perfect world with perfectly rigid, cubical cows, and passengers that are never late you can make flow algo's that reliably hit near 'global maximums'. Things do not change dramatically if you feed the problem to some ML algo.

The flow problem for airlines is in bigger part a problem how to make decisions given near random occurrence of things like sudden worsening of the weather, traffic redirection due to something happening on the runway/airport, fueling trucks being late, and of course the need to maximize revenue per seat. The best decision making algorithm here would not be the one that maximizes throughput, but one that can keep routes more or less consistent given all those changes.

Here, I see that even if you had a near-perfect routing algorithm, and a transportation system with no human driven cars, the whole routing map may have to be changed completely in response minuscule externalities like somebody jaywalking. This class of problems is altogether different than any kind of optimization.

Ironically, the logic class most adapted for solving such problems is the long forgotten fuzzy logic.

You know how solving 90% of some problem takes 90% of the time, and solving the remaining 10% takes another 90%? I have a feeling that with self-driving cars we are not close to 90% yet, not to mention the last 10%. There is a huge difference between "hey, this prototype kind of works" and polished product ready for consumption.
I'm gonna be honest, I have no idea what you are trying to say here.
All kinds of cool inventions never (yet) got past one-offs. Jetpacks, flying cars, exoskeletons (done in the 60's). If the price, safety, usability, etc. aren't all perfect, it can completely fail. Even electric cars got almost nowhere so far.
He makes the same comments about elder care and robots. This will happen to some degree with or without strong AI. I don't think the author realizes that some tasks, like network optimization don't require anything close to strong AI
The thing about self driving cars is the point at which they become generally viable is defined by human social issues.

We could have self-driving cars today. Heck, we could have had them in the 90s. In limited scenarios, and with some small infrastructure investment.

I don't personally view it as a super-hard technical problem. It is a much much harder social problem.

You don't think self-driving is a hard technical problem? Come on man, that's obvious BS.
I think by "hard" he means "might not achieve it". But they actually already exist, so they're automatically possible.
The existing systems could not be defined as self-driving. And the idea that we could have had them in the 90s is ludicrous.
We did have them in the 90s:

http://people.idsia.ch/~juergen/robotcars.html

http://www.roboticstrends.com/article/back_to_the_future_aut...

You seem to have a personal definition of self-driving. What I was saying was to have self driving cars, today you need to:

a) resolve the social issues around self driving cars.

b) install a small amount of infrastructure to enable self-driving cars/standardize roads.

> But... we will

that is a belief. not a fact.