3 ms·
It seems like AI has become more capable but has stayed in the realm of "dumb AI" (not incapable just not human-like). Indeed, by no longer aiming at duplicatin
by joe_the_user 12y ago
It seems like AI has become more capable but has stayed in the realm of "dumb AI" (not incapable just not human-like). Indeed, by no longer aiming at duplicating human consciousness, machine learning has found more and more niches.
But this doesn't mean that right now, the corporations and governments that are deploying this AI and indeed displacing human choice are "provably aligned with human values". Indeed, it almost sound like there's this hope that humans will build machines that treat people right after humans themselves have failed terribly in treating each right. I'm skeptical that this could work.
- Retra 12y agoYou need to know _why_ humans don't treat each other right. Human failure doesn't imply machine failure; that's one of the big reasons we use machines. If a human might fail to add 10,000 numbers properly, this certainly doesn't mean that a machine must fail as well. Some humans can get really good at treating each other right in a 1-on-1 scenario. It's when you scale up to a whole society that things get stupid.
- landryraccoon 12y agoHuman beings can't even agree on a definition of "treat each other right", let alone actually implement it. What makes you think we can program an AI to do it?
- hedgew 12y agoWe have pretty good ideas on how humans or other intelligent agents could "treat each other right". Philosophers, and people who study game theory in particular, have many answers to this problem. The prisoner's dilemma is a simple example of why treating everyone right is challenging.
- Retra 12y ago1. Humans don't have to agree to something to program it. 2. What makes you think a computer can't do something just because you don't know how to do it? Maybe someone else does.
- evanmoran 12y agoI think it's much easier to get 100 people to agree to high moral ideals then to try to get everyone at once. The reason this is possible is only the programmers have to agree.
- eli_gottlieb 12y ago>Human beings can't even agree on a definition of "treat each other right", let alone actually implement it. If you ask 1000 humans to draw a picture or write down a description of things that are Good and Right, they will all draw/write more-or-less the same thing. You could say that, effectively, you would be getting 1000 samples from something like one, or maybe two or three, distributions. Once you'd pinpointed the underlying psychology, learning the model(s) from the data would be easy. Of course, when you start asking those same 1000 humans about the meta-level theory that they believe underlies their picture/description, they will all give ridiculously divergent answers. This is because the overwhelming majority of them are assuming that some kind of elaborate metaphysical theory is required to what basic moral psychology already does, and they also assume that, even if they've been shown the moral psychology, that can't be the real answer (ie: they believe the fact-value distinction is metaphysical). Luckily, a well-built artificial agent doesn't give half a damn about bloody metaphysics! Or I can just pick something I happen to like, and tell the rest of the human race to go jump in a lake.
- dorfsmay 12y agoThis is a post vs pre singularity problem. Post-singularity, the bots won't care about our rules of ethics, rules of engagement etc... Definitely something to worry about. Before we get there though, we have the issue of the dumb bots giving exponential powers to the few humans in charge. Today military and police personnel can refuse to execute an order that is wrong. Dumb bots won't.
- joe_the_user 12y agoToday military and police personnel can refuse to execute an order that is wrong. Maybe, though we see more military and police adding in extra-behaviors that seem rather wrong. A dumb bot merely enforcing a given law might actually more desirable than present-day cops who apparently gratuitously murder civilians on a regular basis.
- ridgeguy 12y agoSo long as (human, artificial) intelligences operate within an evolutionary dynamic, the definitions of what it means to "treat people right" (or even who's included in the class: people) will be embodied in the evolutionary winners. If it happens that human-friendly values degrade fitness in the AI arena, the winners won't have retained or evolved those values. If the values are evolution-neutral, well, perhaps. Evolution requires only variance, selection, and inheritance. I don't know how we'll develop useful ("strong") AI in a setting that excludes AI evolution. I think this means we can't guarantee a human-benign outcome for strong AI.
- joe_the_user 12y agoEvolution is a theory of how biological systems change. Even in regard to biological systems, it's not a theory of what's better but simply what happens. However, outside biological systems, it's not really applicable at all. You can't say a rock, a broom or a computer evolve in the standard sense of Darwinian evolution - only biological systems have the tendency to reproduce themselves. Computers don't reproduce themselves, they are simply produced by us. If an AI is constructed, that will be what happens - human choice (or maybe chance) will be the agent, not biological evolution.
- ridgeguy 12y agoI think evolution is more widely applicable than just to biological systems. To date, we've only studied its operation in biosystems, which, given our status as biological organisms, seems entirely reasonable. But I believe many of the exclusionary limits we put on nonbiological systems (especially computers) increasingly need to be qualified with the word "yet" - as in, "Computers don't reproduce themselves..." (yet). Evolution is a useful, powerful tool even in nonbiological systems. As a limited example, I'd offer the success of 'genetic' algorithms applied to tasks like antenna design (exemplary paper and pdf: http://alglobus.net/NASAwork/papers/Space2006Antenna.pdf http://alglobus.net/NASAwork/papers/Space2006Antenna.pdf). Where I work, we use these methods to find solutions for multiparametric optimization problems involving mechanical dynamic systems. The algorithms first generate a few thousand candidate solver variants, then pick the top ten or so best performers. These are cross-bred, and then seed the next generation of a few thousand candidates. Rinse and repeat, and within surprisingly few generations, we have an algorithm that does what we need. Although humans wrote these algos, they run in a nonbiological context - but since there's variation, selection, and inheritance, they evolve. When I see this behavior running in silico to great effect, it makes me question our limitation of evolution to biological systems. Granted, it's a primitive example, but if I'd dropped by Australia 3.5B years ago, I might have regarded the stromatolite builders as similarly primitive. On a grander scale, see Lee Smolin's concepts of cosmological natural selection (http://en.wikipedia.org/wiki/Lee_Smolin#Cosmological_natural_selection http://en.wikipedia.org/wiki/Lee_Smolin#Cosmological_natural...). And finally - around 13.5B years ago, there was only hydrogen, helium, and a bit of lithium. Today, we're having this conversation. Although we don't (yet) know the full story of how that worked, to me that's an existence proof that nonbiological matter can evolve intelligence. If computational systems incorporate evolution, I don't see a barrier to AI arising along pathways similar to those that led to us. We'll see.
- ableal 12y ago> machines that treat people right You might come to regret that: http://en.wikipedia.org/wiki/With_Folded_Hands http://en.wikipedia.org/wiki/With_Folded_Hands