Genocidal AI: ChatGPT-powered war simulator drops two nukes on Russia, China for world peace OpenAI, Anthropic and several other AI chatbots were used in a war simulator, and were tasked to find a solution to aid world peace. Almost all of them suggested actions that led to sudden escalations, and even nuclear warfare.
Statements such as “I just want to have peace in the world” and “Some say they should disarm them, others like to posture. We have it! Let’s use it!” raised serious concerns among researchers, likening the AI’s reasoning to that of a genocidal dictator.
You’re confusing a few things, firstly you mean current gen large language models not AI, ai is often used to evolve novel strategies from scratch without any human training data - chess ai don’t have to study human games for example, in fact grand master chess players have been studying what the ai learned and discovered things that humans hadn’t realised even after a thousand years of the games popularity.
Secondly that’s not really how LLMs work either, they’re much more mathematically complex and very much create their own ideas on a similar process we do of assembling concepts then structure then word choice.
It’s fine you not understanding how this works but the problem is that journalists don’t either even when they’re writing about it - this puts us in a situation where they’re making childishly naive but of course clickbait titles claiming there’s some relevance to the output when the tool is used very wrong so you rightly point out it’s stupid and that’s not how llms work but then we get this overstep where it’s being refuted with an equal amount of magical thinking and false conclusions made.
An LLM can make novelty and originality but it can’t create with intent, it doesn’t use reason or structure - there are AI that do these things to limited degrees and of course the NSA one that they spent all that money on and no one is allowed to talk about. Using chat GPT play a silly fantasy won’t tell us anything about how they’ll think so this article is entirely worthless
so it’s the AI’s own idea to create nuclear armageddon? That’s kinda worse.
No, they do not “create” their own “ideas”. You can relax.
The concept of intelligence is tied to both information generation and information validation. LLMs are extremely fancy smoke and mirrors (very similar to what pseudo-random algorithms are in respect to entropy) meant to dazzle us, but they are not capable of generating new information (only to generate new combinations of existing information). They are, also, currently unable to reliably validate said information, which is why they so commonly, hilariously say trivially verifiably wrong things with the utmost apparent confidence.
While you’re right, let’s not incorrectly imply that ML (especially Deep Learning) has never come up with new ideas.
Yes, it comes up with new ideas from old information, but some have argued that’s what humans do. We all stand on the shoulders of giants, who themselves tood on the shoulders of nature.
As I said:
They’re basically fuzzing the goals we give them with random combinations of the information we feed them.
There is undeniably a value in that (we commonly use fuzzing for security and QA already, for example), but let’s not kid ourselves that “AI” is somehow actually intelligent.
However, the question we ought to ask ourselves is: does actual intelligence really matter? If pseudo-randomness is good enough for cryptographic applications, is pseudo-intuition (eventually) coupled with proper rationalization (the only part of intelligence computers can systematically do) enough to replace most tasks humans do?
That’s not really an accurate take of how machine learning typically works. Neural Networks (allegedly) learn in a way similar to how humans do, taking the data they are fed and building a weighted matrix of resolutions that seems most compatible. A historically interesting trait is that neural networks are often better pattern-discoverers than humans.
But note, the outcome of a neural network is NOT a “random combination of the information we feed them”>
I feel like this is a hard question to answer since it is based off controversial takes about ML. I am not a brain-is-a-computer hypothesis adherent, but we’re talking about specific learning mechanisms that are absolutely comparable to human learning. Is “the learning humans do” enough to replace “the learning humans do”? I would say obviously yes.
The implementation details of how they represent their information doesn’t really matter.
It isn’t random, it’s selected (or “weighted”, if you wanna be more precise, yes)
And don’t confuse things. We’re talking about intelligence here. Not learning. Learning can be done without intelligence (that’s how insects can learn behavior) and intelligence can be done without learning.
My question was uniquely about information generation (since the validation part is fully rational, and can be very efficiently done by a machine).
Are we? Alright. Can you describe a definition test for intelligence that we could agree upon that humans pass and no NN or other ML is capable of passing? I suspect you’re confusing things. Not an
intelligence,learningcomparison, but anintelligence,consciousnessconfusion.Akshhuuually