What Do You Think about Machines That Think? | Hacker News Reader

What Do You Think about Machines That Think?

edge.org

46 pointsprivong45 comments

My first thought was to make a hard rule that it must obey the laws in the location it's in. In the US, that would be federal, state, and the various local laws of your city or township or whatever.

And then I laughed at the mental image of a robot stuck in a deadlock due to all the mutually exclusive laws we have. (Of course, in Hollywood, replace "deadlock" with "exterminating all humans".)

Even attempting to implement such a rule could have benefits. We could end up with formalized proofs of how those are are objectively contradictory, which in turn could be factored into the design of future laws as a preventative measure. Imagine "test-driven legislation," if you will.

> test-driven legislation

I had a conversation with my brother, an attorney, about this at one point. I vaguely recall him asserting there have been some steps in this direction, but the lawyers always give up.

Please make this a thing!

IANAL, but you'll find that laws tend to be nuanced and written with human judgment in mind. Such efforts are at cross purposes with formalization into a software system.

Closing that gap is the point of the exercise. :)

Attempting to adhere to two conflicting laws results in undefined behavior. And as we all know, undefined behavior can mean anything, including exterminating all humans.

    switch(getJurisdictionalPrecedence(&law)){
       
       case FEDERAL: {
         setPrimeDirective(law[FEDERAL]);
       } break;

       case STATE: {
         setPrimeDirective(law[STATE]);
       } break;

       case CITY: {
         setPrimeDirective(law[CITY]);
       } break;

       default: {
         setPrimeDirective(this->ExterminateAllHumans());
       } break;
   };

You forgot to add this case before the rest:

  case PHYSICS: {
    setPrimeDirective(law[PHYSICS]);
  } break /* at your peril */;

That made me LOL. This might actually make for an interesting way (via A.I.) to find loopholes in law, seeing law as a programmatic way to determine acceptable societal PID's.

We will build very smart machines before we build machines that think. Those very smart machines will probably change the world the way the internet did. Imagine having smart machines than can farm, build houses, clean up pollution, etc. Thinking machines are too far off to know what to expect. We need to first consider what an increasingly automated world would mean. In particular, what will it mean for employment and average income?

We'll figure something out when we get there.

What I found quite annoying lately (more than usual) is that there is a lot of "what if machines, terminator, bla bla bla death! destruction!".

I'm not sure if it's the recent developments with the deep neural nets, maybe it is.. Maybe it's the high profile neo-luddites like Hawkins or Musk. I wish I could have a megaphone and shout: AI that would be classified as having comparable human intelligence is way too far and there are a lot of pieces missing. Just because we can identify with 95.4% accuracy some MNIST dataset or 200 species of dogs, or some guys playing baseball, it's still a long way to getting to a working AGI.

But instead on focusing on doing the actual work that would get us to these machine 'thugs' or at least promoting more research in that direction most prefer to sit in their armchair and argue to the death the hypothetical T-1000.

Thankfully there are smart people out there doing real work and getting results so that is reassuring.

My problem with these extremists is that they produce hype. Hype that got us the AI research winter last time.

Musk, Hawking, et al. are not talking about deep neural nets that do well on MNIST, but what will be possible 20-50 years down the road. There are very real concerns about AI and it's annoying to see them just dismissed by calling people "neo-luddites".

>But instead on focusing on doing the actual work that would get us to these machine 'thugs' or at least promoting more research in that direction most prefer to sit in their armchair and argue to the death the hypothetical T-1000

There has been more investment in AI in the last 5 years than in the 50 years before it. Even Musk invested millions to AI research.

Spending time and effort on AI "20-50 years down the road" is a waste because:

(a) statistically based, but obviously not (or not obviously) intelligent systems are going to change the legal and practical landscape well before Ultron gets up and starts walking around, and

(b) the kinds of things they're worried about are almost certainly going to appear silly when the real thing does roll around.

The moral responsibility for a bot that orders drugs certainly belongs to the person who set up the bot to order from the "black market"; for a drone that kills the wrong people, it belongs primarily to the officer that ordered the mission knowing that the statistical software wasn't perfect. In my opinion. The supreme court may differ.

Also in my opinion, the dangers of "singularity" AI are overblown due to a soft, fuzzy, but very real upper limit on "intelligence".

I retract my previous statement. Musk is not a luddite. He's a clever businessman.

We are already gearing them toward good... "A Search Engine that Thinks, Almost" http://newscenter.lbl.gov/2005/03/31/a-search-engine-that-th...

Deleted comment

And you think I don't want to live on my own platform in some giant habitat with drones that do everything for me. Where disease and famine is eradicated, where I can change my body however I want and I have no concept of money? Sure! Just as much as the other guy, but to get there we need some hard science, funding and steady progress.

AI is one of the hardest problems facing us so doing things we so notoriously bad at like estimating the difficulty of a such problems will not get us any results.

I encourage skepticism and careful research as opposed to ignorance and fear-mongering.

On that note why was not Lecun's FAIR project not on front-page HN? Instead we have this shit... Is there any hope that we'll grow up?

The machine (software program) lacks sufficient awareness beyond mechanical execution of the programmer's own will; therefore, it holds infinitesimal responsibility for its actions. In order to prevent further violations of law, it should be decommissioned. The programmer should be fined equivalent to the amount of illegal purchases the bot made in addition to $1K for each offending purchase.

As for other machines in the future (androids), it seems obvious enough to me that - insofar as they are intelligent and thus responsible (IQ testing is commonly used in US judicial system, with the cutoff being an IQ of 70 or above) - they can be tried like any human being.

Culpability goes all the way to the top. So the top ranking official who orders use of the machine is to blame, which could very well be the president. And it goes deeper than that, because the society that elected the president endorses its use if it doesn't recall him or her. Since recalls are rare, collateral damage from thinking machines is basically inevitable and I don't think most people have any idea what is coming.

That said, I expect that AI is going to transcend the way people solve problems to such a degree that it may even alleviate the underlying causes of war like artificial scarcity and ignorant intolerance.

The article has nothing to do with thinking. But in the scope of the article, it's an interesting legal conundrum. I expect the end result is that the class of software which can perform illegal actions autonomously will have to be brought under warranty and the creators held responsible.

"Acme Inc. Warrants that under Factory Conditions and Standard Programming, Software Agent R2-D2 will not conduct illegal actions".

This is not out of the range of possibility and even feasibility today, but it will require genuine engineering mentalities to create such systems.

The same way we deal with other entities: rats, viruses, hurricanes. If they harm us, we don't "punish" them, we just seek a way to prevent them harming us.

Why was the title changed to the Edge.org question of the year?

We can give them negative reward (decrement some counter in the machine memory). The machine will want to avoid that!

Machines killing Machines is BETTER than Humans killing Humans

Sure, but machines killing humans scales better than humans killing humans.

What do machines think about machines that think?

And what they think about you?

Just follow Asimov's laws of robotics:

0) A robot may not harm humanity, or, by inaction, allow humanity to come to harm.

1) A robot may not injure a human being or, through inaction, allow a human being to come to harm.

2) A robot must obey the orders given it by human beings, except where such orders would conflict with the First Law.

3) A robot must protect its own existence as long as such protection does not conflict with the First or Second Law.

If corporations are people, then robots are people too.

Corporations pretty much only understand monetary fines, and robots will only understand some form of disconnecting the cord.

Similarly, corporations have a corporate veil to shield their owners form liability, assuming their business finances are separated from personal. Robot designers would have limited liability assuming their personal decision making was separated from the algorithmic reasoning.

And we've managed to pass the CAN-SPAM Act, so either we should restore robot 1st Amendment rights, or otherwise overturn Citizens United.

If corporations are people, then robots are people too.

Corporations have legal personhood because they're an association of people. I don't see how this applies to robots. If you do, I don't want to see your robots!

Corporations can be wholly owned by other corporations. And they often have an obligation to seek to maximize profits above all other concerns. How's that an association of people?

They're treated as people because there was a similar conundrum as to how to apply laws to corporations. Justices decided to start applying laws as though they're people. It's the same situation Schneier is describing with robots.

>Corporations can be wholly owned by other corporations. [...] How's that an association of people?

If you follow that tread long enough, you will eventually find people running the show. You will not find robots, sentient trees, or rocks sitting around a board table.

Hm... now you have me wondering if I could slip a circular chain of ownership past the chambers of commerce here.

I'm pretty sure you could bootstrap this:

Person 'X' incorporates company 'A' and receives the shares.

Company 'A' incorporates company 'B'.

Company 'B' makes some money.

Company 'B' buys the shares of company 'A' from person 'X'.

Presto, no more people involved, just two companies owning each other. It might even be possible to have company 'A' buy the shares of company 'A' directly of the person that incorporated it, so it would own itself.

You'd have to fiddle with the valuation during the transaction because you're using the company capital to do the purchase, essentially ending with a valuation (temporarily) of zero when the transaction has completed after which the company would have to create some more income to become liquid again.

Tricky. Very much tempted to try this.

"You'd have to fiddle with the valuation during the transaction” Just have company B use some leverage aka get a loan to buy A. That way it's valuation does not need to change pre or post purchase.

PS: Wondering if this has happened somewhere such as Japan where they have some very convoluted ownership diagrams.

Good point!

I wonder how you'd go about the usual annual shareholder meetings and things like that. Without a real person to represent even a single one of the shares it would be pretty hard to get a vote on anything. Board of directors would be a bit problematic, hiring a CEO could be done but who would do the hiring?

This is not uncommon situation when some venture fund owns a large share in a company they just appoint someone to sit on the board. The appointee need not actually own shares they just represent shares, so company A sends Sam to be on B's board and company B sends Ted to sit on A's board.

The funny thing taxes would just end up eating up any profit as B sends say 70% of their profit to A as a dividend which sends 70% of 70% to B as a dividend and repeat.

You don't have to pay out dividends, and even if you did in 100% ownership situations dividends between companies are not taxed at all (at least not here).

I don't think you need all those steps. Just create the corporation, then donate all of your shares to the corporation. It now owns itself.

But who makes decisions for it and signs the paper work? I don't think there is any way around having a human in the loop somewhere for legal purposes, and that would defeat the purpose.

It would be fairly easy to create a bot that makes random corporations. Legally I don't think there is anything that would separate them from actual corporations.

Granted, their corporate officers would presumably need to be people. But, HFT demonstrates bot's can make legally binding decisions so that might be a grey area.

Someone is still responsible for those bots.

What if the bot creator dies before the bot makes a decision. We could even go so far as to have a cosmic ray flip one or more bits in memory that causes this to be released into the wild.

Then the sole shareholder of the company dies. I'm sure there's plenty of case law regarding this situation.

Sure, but shareholders are rarely responsible for a companies actions: http://en.wikipedia.org/wiki/Piercing_the_corporate_veil

Anyway, my point is if bot's can create and run companies then at some point they could gain the use of corporate personhood.

And they often have an obligation to seek to maximize profits above all other concerns.

Only if the people who form that association demand it. There's no legal obligation - that's a common myth.

ASSIMILATE

Reply on news.ycombinator.com