An LLM fundamentally knows nothing about the world, only how to put together language in a statistically convincing manner. It makes no empirical observations about the world, only about how words connect to different words, which is definitionally not knowledge about anything other than language. It does not matter how powerful this technique becomes, it will always have this limitation.
The real use cases for LLM or machine learning are pattern recognition tasks that would be impractical for a person to manually do. If the task has a narrowly defined goal with predictable error rates that can be accounted for, you can parse impossible amounts of data without making any leaps in reasoning. All uncertainties are accounted for. Machine learning has been used this way in research for far longer than most people realize, and it has enabled analysis that would not have been possible otherwise.
Asking an LLM to give you factual information or even asking it to summarize sources is not a narrowly defined task with predictable inaccuracies. There is no real way to know the error rates of questions you ask an LLM, and even wording the question slightly differently can result in a different outcome. All you can do with an LLM is ask it for sources, using it as a supplement for a search engine(what I’ve always thought they should use the tech for), but the summaries it gives you are basically a waste of processing power.
The key to all of this is critical thinking, something LLM use actively atrophies. It has no consistent viewpoints and does not think like a person. It cannot gather information firsthand and can only confide with others for all its information about the world. It does not matter how good machine learning gets; at the end of the day it is either reliant on the information of others, or just spewing bullshit that sounds like what others would say. Just cut out the middle man and interface directly with people
An LLM fundamentally knows nothing about the world, only how to put together language in a statistically convincing manner. It makes no empirical observations about the world, only about how words connect to different words, which is definitionally not knowledge about anything other than language. It does not matter how powerful this technique becomes, it will always have this limitation.
The real use cases for LLM or machine learning are pattern recognition tasks that would be impractical for a person to manually do. If the task has a narrowly defined goal with predictable error rates that can be accounted for, you can parse impossible amounts of data without making any leaps in reasoning. All uncertainties are accounted for. Machine learning has been used this way in research for far longer than most people realize, and it has enabled analysis that would not have been possible otherwise.
Asking an LLM to give you factual information or even asking it to summarize sources is not a narrowly defined task with predictable inaccuracies. There is no real way to know the error rates of questions you ask an LLM, and even wording the question slightly differently can result in a different outcome. All you can do with an LLM is ask it for sources, using it as a supplement for a search engine(what I’ve always thought they should use the tech for), but the summaries it gives you are basically a waste of processing power.
The key to all of this is critical thinking, something LLM use actively atrophies. It has no consistent viewpoints and does not think like a person. It cannot gather information firsthand and can only confide with others for all its information about the world. It does not matter how good machine learning gets; at the end of the day it is either reliant on the information of others, or just spewing bullshit that sounds like what others would say. Just cut out the middle man and interface directly with people