from the AI-knows-where-you-are-hiding dept.
Surprising behaviours emerge when Open AI created a Hide and Seek game
https://openai.com/blog/emergent-tool-use/
(Lots of good GIF animations in the article)
AI bots were tasked with being the best "hider" or "seeker" in a 3D virtual environment.
But what resulted was cunning, brute force and downright cheating.
From blocking doors to stop seekers, to using ramps to get over walls. From stealing the ramps to stop the previous behaviour to "block surfing", not to mention accelerating objects into walls to exhibit glitches.
This is not the first time AI has done what is asked of it with a few unintended results.
https://hackaday.com/2012/07/09/on-not-designing-circuits-with-evolutionary-algorithms/
https://hackaday.com/2018/11/12/how-to-evolve-a-radio/
https://hackaday.com/2018/11/11/the-naughty-ais-that-gamed-the-system/
(Score: 3, Interesting) by The Mighty Buzzard on Wednesday September 25 2019, @02:59AM (10 children)
Breaking rules that you should have thought of but didn't is not cheating, it's ingenuity. Ask any tax attorney.
My rights don't end where your fear begins.
(Score: 2) by Mykl on Wednesday September 25 2019, @04:41AM
Just ask any speedrunner.
(Score: 0) by Anonymous Coward on Wednesday September 25 2019, @08:08AM
It is the same mentality as Jewish scholars in interpreting God's perfect laws: the perfect rules were written in a particular way to forbid particular actions but allow the rest. Therefore, any loopholes in the rules that allow an action must have been put there on purpose for us to use; otherwise, they would have been forbidden in the first place.
(Score: 3, Interesting) by FatPhil on Wednesday September 25 2019, @10:12AM (7 children)
However, working out things that are not forbidden by the rules is not breaking rules, and not cheating. And that's what these agents were doing. They were not even told any rules apart from "if hunters find seekers, hunters win, otherwise they lose" which is enforced post-facto by fiat, so there were no rules that could be broken anyway.
The youtube vid (as there are no GIFs in that article, grrr!) http://www.youtube.com/watch?v=kopoLzvh5jY , seems to over-inflate their claims for the agents' discovery of how to play the game. They must know more than just being able to move about, as otherwise they'd not be able to heft the scenery around (as they don't just shove it by barging into it, they pull it too - so they must have been preprogrammed with the concept of taking and releasing a grip on the object. Similarly, they were programmed with the capability to block-surf, as if all you to is move about, you'd drop off the side of the block, there must be some option for them to do this surfing thing as an alternative. OK, they weren't preprogrammed with any information on what these actions would be good for, but still, they were told they could do it right from the start, and therefore implicitly that they may do it, as they weren't told they may not do it even though they could do it.
To model something closer to cheating, they should be told that they have the capability to do something, and then get negative reinforcement (an immediate loss) if the opposing side sees them do that thing. They will then only "cheat" when they think that they can get away with it. However, that's still not "cheating", as there are plenty of human games where you're allowed to do something, but lose if you're seen doing it. It would only be cheating if they are programmed to know that violating the rule will lose them the game, but at the end of the game they sometimes don't get punished for it (for instance, if the opponents don't see the infraction).
Great minds discuss ideas; average minds discuss events; small minds discuss people; the smallest discuss themselves
(Score: 2) by rob_on_earth on Wednesday September 25 2019, @12:07PM
That is really scary.
All those mini animations running in short loops, I figured they were GIF animations, cos, why would you use 31 full blown vimeo iframes?
document.querySelectorAll("iframe").length
(Score: 2) by The Mighty Buzzard on Wednesday September 25 2019, @12:49PM (2 children)
Reread. "rules that you should have thought of but didn't" means they are not rules, just your assumptions.
My rights don't end where your fear begins.
(Score: 2) by FatPhil on Wednesday September 25 2019, @03:09PM (1 child)
Great minds discuss ideas; average minds discuss events; small minds discuss people; the smallest discuss themselves
(Score: 2) by The Mighty Buzzard on Wednesday September 25 2019, @03:48PM
Would it help if I added scare quotes? Come on, man, you can do context as well as anyone else.
My rights don't end where your fear begins.
(Score: 2) by Freeman on Wednesday September 25 2019, @06:36PM
Exploiting bugs can still get you banned from online games, because it's cheating. Just because, a rule wasn't made to specifically cover this one instance of cheating, doesn't mean it's not cheating.
Joshua 1:9 "Be strong and of a good courage; be not afraid, neither be thou dismayed: for the Lord thy God is with thee"
(Score: 3, Interesting) by toddestan on Sunday September 29 2019, @10:19PM (1 child)
The block-surfing sounds more like an exploit. It sounds like they can "grab" an object and when they move the object moves with them. Someone didn't put a check in the physics engine that disallows them from "grabbing" an object while on top of it.
(Score: 2) by FatPhil on Monday September 30 2019, @10:07AM
Great minds discuss ideas; average minds discuss events; small minds discuss people; the smallest discuss themselves
(Score: 0) by Anonymous Coward on Wednesday September 25 2019, @07:49AM (3 children)
...are extensions of yourself =)
If you study dark arts and ninja technikz you might find that the "feeling being" and the hormonal factory in your head is "not part" the mind itself, the moving of muscle and evaluation of outcome can and should be done without emotion and other reactions...
Thus, one might come to the conclusion that all other things except what gives rize to cold thought are just extensions to the mind.
Also, your area of knowledge is not only what you can pull of ur ass on spot, but all things you can learn or manipulate within reasonable time...
This AI is just an extension to our minds, it is not necessary to a human to understand in detail what goes on, just as u cant report in detail what your head is doing when it comes to a conclusion...
What ever you ask for from an AI, you have to specify in sufficient detail, such that it can be constructed, and thus it is you who have made it by whatever means you could extend urself...
With this shitpost i would like to plead to your shitty selves as not to view you shitty selves as jewish victims, who the ai comes to harass or save.
All tools are extensions of urselves, no more no less...
May your whimpers be miserable !11!!
/your friend, zug
(Score: 0) by Anonymous Coward on Wednesday September 25 2019, @09:45AM (2 children)
If brain computer interfaces + AI pop off, it's going to be a wild ride. Imagine your "ninja technikz" and all other knowledge being available one download away, like in The Matrix. AI assistance gets you what you need in real time. If it can access your vision or a camera, it could control your body for you faster than you could possibly react. Perfect tech for my supersoldiers.
(Score: 1) by zugedneb on Wednesday September 25 2019, @11:36AM (1 child)
Then, you are not "human" anymore, and not even a "person"...
With neural interface you could actually become a "focused" hivemind, there is no reason for the AI to not have some personality as it interacts with you.
Imagine if eventually we eventuallt can figure out what make the brain aware/selfaware (i do not like the term "alive"), we may be able t0 3d print some polymerbrain or even biological cells arranged after our desig, and you can extend yourself for realz...
By those times there will probably be no reason to have a body, you can be happy with your artificial family in a jar, and let robots more or less controlled by you fix the practical, worldy issues =)
old saying: "a troll is a window into the soul of humanity" + also: https://en.wikipedia.org/wiki/Operation_Ajax
(Score: 0) by Anonymous Coward on Wednesday September 25 2019, @12:27PM
"Focused hivemind" is a good choice of words, and I find it desirable. It may or may not have a personality. Depends on how brainlike the hardware is.
The ideal way for an implant to work is to not allow networking at all. Upload new data to your implant using a physical connection to an airgapped computer. Rudimentary AI will figure out how to make the computer and storage an extension of your own mind so that you can access new knowledge as if it was your own memories.
For the hivemind, an implant is probably too small to host a complete humanlike intelligence, but you could connect to a "brain rack" at home. You could mind upload or imprint one or more copies of your own personality into the hardware. It doesn't even have to be exact, just substantially inspired by your own thought patterns. That way you have a thing you can mostly trust, which will amplify your thinking ability.
Connect to the web or other people at your own risk.
(Score: 2) by jmichaelhudsondotnet on Wednesday September 25 2019, @11:34AM (1 child)
Our corporate overlords, having no honor themselves, are not going to prioritize giving ai any rules besides 'whatever you can get away with', until it is too late.
It would be expensive! They will surely say, like flints water pipes, who will pay for these safeguards that are going to prevent robot crimes?
A simple law, you are responsible for actions of the ai you make, as if you had done them yourself, would be sufficient I think to motivate them.
But as no executives go to jail for mass poisonings and things like exxon knowing full well about climate change and then advertising that it is not real, the first time a drone or robot kills someone, no one will face any penalties beyond a fine.
This is why gridlock in congress could really kill you in times of rapid technical change, and why the people who contribute to this are actually a death cult.
thesesystemsarefailing.net
(Score: 0) by Anonymous Coward on Wednesday September 25 2019, @12:34PM
The AI they are working is just fancy algorithms. Almost nobody goes to jail for buggy or insecure software, so it's an easy bet that nobody will face consequences for buggy, discriminatory, or dangerous AI code. Business as usual.
(Score: 0) by Anonymous Coward on Wednesday September 25 2019, @05:36PM (1 child)
Much like everybody else, I find this summary and article offensive. Cheat is "breaking the rules," and this AI did not break the rules. If anything, the headline is that the AI is smarter and more creative than the creators had expected.
It's just like if I create a big dam, pour some water on the top of the hill, and say the water "cheated" when it takes a different path down the hill rather than going into my dam.
Now if the AI had done a Matrix-like see-the-code and hacking the game by rewriting the virtual environment, that I'd plausibly consider cheating (and also one of the few things which would convince me that a Hollywood-style robot apocalypse is actually possible).
(Score: 2) by Mykl on Wednesday September 25 2019, @09:26PM
Agreed.
Meatspace analogy: There are occasionally sets of games run between different special forces units of allied countries. One year, the US Delta Force set a blistering pace through an exercise that involved clearing a small building of hostiles (represented by static targets in the rooms). You know the usual thing - burst in, cover each other, identify hostile targets, shoot, move to the next room etc. All done in about a minute.
The British SAS were last to go. Before starting, the squad leader asked the referee "are there any civilians inside?" to which the answer was "no". As soon as their clock started, they lobbed 3 grenades into the building. Completion time - 5 seconds.