Idk why you are being downvoted. It is a real possiblity, but i just dont think its likely.
reliable and cheap
Because it simply wont ever be those things. LLMs become more expensive the better you try to make them, exponentially so. We will likely never see LLMs that are significantly better than what we have today. Basically all written human texts on the internet have been used for LLM training. Any further gains can only come from optimizing training algorithms and there hasnt been any real progress there for years now.
Yes, it will take a substantial change in LLM architecture to make them viable for general use. Like you said, a couple orders of magnitude improvement in training would be the biggest help. Though I’m not sure that’s even possible.
The internet was successful because it was predictable and reliable and filled a previously unserved purpose. LLMs could possibly serve a purpose creating text that doesn’t need to be correct or accurate, but it won’t replace anything that requires any level of accuracy or accountability because the design is inherently flawed due to being a prediction algorithm.
There are uses for pattern matching AI that will continue to be tools used by humans to help focus efforts in massively complex work, but they won’t ever be foundational replacements.
Idk why you are being downvoted. It is a real possiblity, but i just dont think its likely.
Because it simply wont ever be those things. LLMs become more expensive the better you try to make them, exponentially so. We will likely never see LLMs that are significantly better than what we have today. Basically all written human texts on the internet have been used for LLM training. Any further gains can only come from optimizing training algorithms and there hasnt been any real progress there for years now.
Yes, it will take a substantial change in LLM architecture to make them viable for general use. Like you said, a couple orders of magnitude improvement in training would be the biggest help. Though I’m not sure that’s even possible.
The internet was successful because it was predictable and reliable and filled a previously unserved purpose. LLMs could possibly serve a purpose creating text that doesn’t need to be correct or accurate, but it won’t replace anything that requires any level of accuracy or accountability because the design is inherently flawed due to being a prediction algorithm.
There are uses for pattern matching AI that will continue to be tools used by humans to help focus efforts in massively complex work, but they won’t ever be foundational replacements.
Oh yeah for sure there are plenty of great uses for machine learning, i would never deny that.