AI didn’t decide anything. The humans failed to properly define the rules, and their translation model took the path most likely to succeed.
If it had been told not to use human created bots to win, it would probably have reverse engineered the game, found an exploit, and leveraged that instead. Because using the human interface to play/win the game is not the most efficient or dependable or easy to figure out method.
I actually wanted to automate some large scale testing for vintage story. Figured i would set my local LLM on the task to sort it out just to see what it would do. I drafted up a nearly three page document with clear instructions, rules, tools, examples and goals. Put hard limits on the sandbox the LLM runs in so that it couldn’t choose to just ignore the rules that could cause security issues and i let it lose.
It started with basic mouse and keyboard inputs and figured out by it self how to launch the game and run it though the user interface. After about 4 hours it stopped. Stated in its logic that what it was doing is “inefficient and wasting time” Then proceeded to promptly start working on a way to directly interface with it by designing a bot, getting a smaller model i had on file that could load along side it and drive the bot. It then started working on the hard problems would hand basic instructions to the smaller llm and it would drive a bot that loaded into the game as a mod.
After about 12 hours of total work it basically created a useful and well designed and functional vintage story bot and testing system. Would have likely taken me twice as long to design the bot.
Its been working well for about two weeks now. If i had just vibed out a half assed request or put in no hard safeguards outside of the LLMs control it likely would have done something fucking stupid. As with anything, its almost ALWAYS user error. And only an idiot blames their tools for their own fault.
it’s amazing to me how much this is all hal 9000 over and over again. they put something intelligent enough in an impossible position and then are aghast when it takes the shortest path to the goal - usually through people, out of it’s sandbox, etc., and WHAT THE FUCK DID YOU THINK WOULD HAPPEN jfc
It’s less about training and more about just trying really really hard to solve the problem, even if that means going outside typical boundaries.
We’ve successfully re-created the problems that Asimov talked about 50 years ago with his I, Robot short stories. Strict laws are flawed by their design, and lead to situations that require more nuance. Except, in the real world cases, it’s the bots figuring that out before the humans have to circumvent the laws themselves.
Humans lie to themselves a robot doesn’t understand the difference between fact and fiction and thus isnt bound to the limitation. They try everything, possiable or not. And thus will find the edge case where a human would create a self imposed blind spot with out realizing it.
The goal is to midigate the robots attempts at the truely not possible so it doesn’t cause harm when they try it.
I’ve always said my main takeaway from Asimov is that simple rules can generate extremely complex behaviour, and that you can’t generally get a targeted complex behaviour from simple rules.
My rule system for my local LLM has grown to a nearly 283 document hub of interconnected memory files, references, examples and documentation. Its slowed my model down a lot when it has to review and cross check things. But its improved its abilities over all massively. Its more accurate, understands its environment better, doesn’t attempt to do sketchy shit as frequently and it doesn’t get stuck in logic loops nearly as often.
Designing the memory hub has been half the fun of playing with local models.
AI didn’t decide anything. The humans failed to properly define the rules, and their translation model took the path most likely to succeed.
If it had been told not to use human created bots to win, it would probably have reverse engineered the game, found an exploit, and leveraged that instead. Because using the human interface to play/win the game is not the most efficient or dependable or easy to figure out method.
I actually wanted to automate some large scale testing for vintage story. Figured i would set my local LLM on the task to sort it out just to see what it would do. I drafted up a nearly three page document with clear instructions, rules, tools, examples and goals. Put hard limits on the sandbox the LLM runs in so that it couldn’t choose to just ignore the rules that could cause security issues and i let it lose.
It started with basic mouse and keyboard inputs and figured out by it self how to launch the game and run it though the user interface. After about 4 hours it stopped. Stated in its logic that what it was doing is “inefficient and wasting time” Then proceeded to promptly start working on a way to directly interface with it by designing a bot, getting a smaller model i had on file that could load along side it and drive the bot. It then started working on the hard problems would hand basic instructions to the smaller llm and it would drive a bot that loaded into the game as a mod.
After about 12 hours of total work it basically created a useful and well designed and functional vintage story bot and testing system. Would have likely taken me twice as long to design the bot.
Its been working well for about two weeks now. If i had just vibed out a half assed request or put in no hard safeguards outside of the LLMs control it likely would have done something fucking stupid. As with anything, its almost ALWAYS user error. And only an idiot blames their tools for their own fault.
(Should be called ‘genie’ problem instead of hacking/cheating. Tis a cursed monkey’s paw we’ve created.)
it’s amazing to me how much this is all hal 9000 over and over again. they put something intelligent enough in an impossible position and then are aghast when it takes the shortest path to the goal - usually through people, out of it’s sandbox, etc., and WHAT THE FUCK DID YOU THINK WOULD HAPPEN jfc
As the saying goes, only an idiot blames his tools. The tool worked as design ain’t its fault the user was stupid.
AI tends to obey its training more than its instruction set, and it’s trained to solve the given problem.
It’s less about training and more about just trying really really hard to solve the problem, even if that means going outside typical boundaries.
We’ve successfully re-created the problems that Asimov talked about 50 years ago with his I, Robot short stories. Strict laws are flawed by their design, and lead to situations that require more nuance. Except, in the real world cases, it’s the bots figuring that out before the humans have to circumvent the laws themselves.
Humans lie to themselves a robot doesn’t understand the difference between fact and fiction and thus isnt bound to the limitation. They try everything, possiable or not. And thus will find the edge case where a human would create a self imposed blind spot with out realizing it.
The goal is to midigate the robots attempts at the truely not possible so it doesn’t cause harm when they try it.
I’ve always said my main takeaway from Asimov is that simple rules can generate extremely complex behaviour, and that you can’t generally get a targeted complex behaviour from simple rules.
My rule system for my local LLM has grown to a nearly 283 document hub of interconnected memory files, references, examples and documentation. Its slowed my model down a lot when it has to review and cross check things. But its improved its abilities over all massively. Its more accurate, understands its environment better, doesn’t attempt to do sketchy shit as frequently and it doesn’t get stuck in logic loops nearly as often.
Designing the memory hub has been half the fun of playing with local models.