4 Experiments Where the AI Outsmarted Its Creators! 🤖

@TwoMinutePapers says:

Agreed. I would go so far as to say there is little reason to think a superintelligence would do anything else than find the simplest loophole to maximize the prescribed objective. Even rudimentary experiments seem to point in this direction. We have to be wary of that.

@michaelemouse1 says:

So AI would be like a genie that grants you all your wishes, exactly as you ask, in a way that catastrophically backfires. This should be the premise of a sci-fi comedy already.

@scno0B1 says:

why would a robot not choose the simplest solution? we can see that a robot does come up with the simplest solutions 😛

@davidwuhrer6704 says:

That is the secret.
The researches who formulated the problem thought there was a box.
They expected the AI to think inside it.
But the AI never knew about the box.
There was no box.
And the AI solved the problem as stated outside it.

@planetary-rendez-vous says:

So you mean humans are conditioned to think inside a box

@milanstevic8424 says:

the error is simply in trying to describe a very simple “box” while not being able to reconstruct what’s actually described. people do this all the time, and this is why good teachers are hard to find.

the box that the AI couldn’t circumvent was the general canvas, or in this case the general physics sandbox with gravity acceleration and a ground constraint. this is the experimental >reality<, along with the clears goals set by the scientists (use muscles and joints to move from A to B, minimize leg contact with the ground). it is the scientists' inability to investigate the problem space and imagine potential solutions that lead them to this issue. sometimes this is near impossible, as someone compared it with the halting problem, but more often than not, it's an issue of not being particularly imaginative. the skill of understanding the problem space is incredibly important in the fields of programming and game design. in other words, in creating very complex but fully interactive state machines, design of which tends to be impossible to grasp with limited human cognition, and therefore has to be explored strategically or systemically. in case of this particular AI problem, this should really be called "the jinn problem" -- or 'be careful what you wish for' -- similar to how Mulder (from the X files), in the episode "Je Souhaite" ( https://en.wikipedia.org/wiki/Je_Souhaite ) wished for a world’s peace only to be introduced with a world without humans. the wish is obviously fulfilled from the standpoint of a local minimum, which is exactly what the AI does, but it can be stated that any intelligent agent would do this. when a human does this in a loosely defined competitive domain — we call that cheating and/or exploiting.

therefore i.e. slavery is a perfect solution for cheap workforce if only you’d expand your “box” or think outside of any ethical values.

hence, the rise of laws* in “civilized” societies implores the existence of basically immature and intrinsically imbalanced rules, trying to punish any unwanted cheats and exploits in the system, to minimize exploitative behavior and further repercussions.

thank god we don’t make games like that, as we typically root out the causes and/or explore better systemic solutions to imbalances.

* obviously this does not include laws that govern rules made for political and regulatory reasons, as these are implemented to achieve something else.

@milanstevic8424 says:

we’d do it too if only there was such a glitch in the system.
actually we exploit the nature for any such glitch we can find.
thankfully the universe is a bit more robust than our software, and energy conservation laws are impossibly hard to circumvent.

@Moreoverover says:

Reminds me of what video game speedrunners do, finding glitches is goal number uno.

@milanstevic8424 says:

@@jetison333 I agree, even though rocks do not think (yet).

But what is a human if not just a thinking emulsion of oil (hydrocarbons) and water? Who are we to exploit anything that wasn’t already made with such a capacity? We are merely discovering that rocks aren’t what we thought they were.

Given additional rules and configurations, everything appears to be capable of supernatural performance, where supernatural = anything that exceeds our prior expectations of nature.

“Any sufficiently advanced technology is indistinguishable from magic”

Which is exactly the point at which we begin to categorize it as extraordinary, instead of supernatural, until it one day just becomes ordinary…

It’s completely inverse, as it’s a process of discovery, thus we’re only getting smarter and more cognizant of our surroundings. But for some reason, we really like to believe we’re becoming gods, as if we’re somehow leaving the rules behind. We’re hacking, we’re incredible… We’re not, we’re just not appreciating the rules for what they truly are.

In my opinion, there is much more to learn if we are ever to become humble masters.

@dancingdoungnut says:

So fucking true