Given that AI is a human construct, that its "intelligence" is borrowed from and input by all too fallible humans, runs through algorithms devised by equally blinkered humans; it is inconceivable we should predict a more useful product than the trash human brains conjure on digital media.
The algorithms modern AI run on are not designed by humans. Deep learning models such as LLMs) are grown rather than programmed. Humans design the learning algorithm, and then it runs automatically and pre-trains the AI on essentially the whole internet. In post-training, models now do a lot of reinforcement learning on automatically verifiable challenges, especially in math and computer science. In some domains, the models essentially train themselves by testing their reasoning directly against reality, no human data needed. We should expect the reliance on human data to continue to decrease as AI systems become better at grading each other in fuzzier domains.
It's tempting to imagine AI as belonging to the bucket "computer program" and apply all the limitations you are aware of from traditional software, but those limitations do not all apply to this new class of system. Human brains are not magic, and while LLMs are missing a few attributes, they are still able to do a shocking amount of what human brains can do. So when you think you have an argument for why AI can't do something humans can do, you have to be extra careful to ensure that the premises of the argument have actually been robustly proven true of LLMs, or at least that the same argument wouldn't also prove that humans can't do it either.
The danger of AI is not in the ways in which it fails, but in the ways in which it is able to succeed beyond human capability, which are frequently increasing in both degree and number.
Given that AI is a human construct, that its "intelligence" is borrowed from and input by all too fallible humans, runs through algorithms devised by equally blinkered humans; it is inconceivable we should predict a more useful product than the trash human brains conjure on digital media.
The algorithms modern AI run on are not designed by humans. Deep learning models such as LLMs) are grown rather than programmed. Humans design the learning algorithm, and then it runs automatically and pre-trains the AI on essentially the whole internet. In post-training, models now do a lot of reinforcement learning on automatically verifiable challenges, especially in math and computer science. In some domains, the models essentially train themselves by testing their reasoning directly against reality, no human data needed. We should expect the reliance on human data to continue to decrease as AI systems become better at grading each other in fuzzier domains.
It's tempting to imagine AI as belonging to the bucket "computer program" and apply all the limitations you are aware of from traditional software, but those limitations do not all apply to this new class of system. Human brains are not magic, and while LLMs are missing a few attributes, they are still able to do a shocking amount of what human brains can do. So when you think you have an argument for why AI can't do something humans can do, you have to be extra careful to ensure that the premises of the argument have actually been robustly proven true of LLMs, or at least that the same argument wouldn't also prove that humans can't do it either.
The danger of AI is not in the ways in which it fails, but in the ways in which it is able to succeed beyond human capability, which are frequently increasing in both degree and number.
AI VERY DANGEROUS FOR PEOPLE
GIGO
"Garbage in, garbage out" isn't true of LLMs. Adding noise to the data can actually increase performance.
https://arxiv.org/abs/2404.18191
GIGO isn't true of humans for the same reason: we are capable of understanding the underlying principles that generate the data.
Ai is not jus loosing people jobs. It's being used to falsely influence people's votes
Thankyou.