3 ms·

> LLMs will tend to spit those out too, purely because the high school version of anything has been represented heavily enough in the training data that it's mo

> LLMs will tend to spit those out too, purely because the high school version of anything has been represented heavily enough in the training data that it's more likely than not to fall out of the huge matrix of numbers! That assumes that the high school version of the subject exists, which is unlikely because I already have the high school version of most subjects that have a high school version. The subjects that I would want to dig into at that level would be something along the lines of chemical engineering, civil engineering, or economics—subjects that I don't yet know very much about but have interest or utility for me. These subjects don't have a widely-taught high school version crafted by humans, and I don't trust that they would have enough training data to produce useful results from an LLM.

fragmede 2y ago
at what point does a well-read high school-level LLM graduate to college? I asked one about Reinforcement Learning, and at first it treated me like the high schooler, but I was able to prod it into giving me answers more suitable for my level. Of course, I don't know what's hallucinated or not, but it satisfied my curiosity enough to be worth my while. I'm not looking to change careers, so getting things 100% right about in the fields of chemical engineering, civil engineering, or economics isn't necessary. I look at it as the same way I think of astrophysics. After reading Steven Hawkins book, I still don't really know astrophysics at all, but I have a good enough model of things. And as they say, all models are wrong, some are useful. If I were a lawyer using these things for work, I'd be insane to trust one at this stage, but the reality is I'm not using my digging into things I don't know about for anything load bearing, but even if I were, I'd still use an LLM to get started. Eg the post didn't state how the author learned anything the name for the dropped letter O, but I can describe a thing and have the LLM give me the name of it. there's an emphasis on getting things totally 100% right does erode trust, but you get a sense for what's could be a hallucination and then check background resources if you get enough experience with the tool.

fragmede 2y ago

at what point does a well-read high school-level LLM graduate to college? I asked one about Reinforcement Learning, and at first it treated me like the high schooler, but I was able to prod it into giving me answers more suitable for my level. Of course, I don't know what's hallucinated or not, but it satisfied my curiosity enough to be worth my while. I'm not looking to change careers, so getting things 100% right about in the fields of chemical engineering, civil engineering, or economics isn't necessary. I look at it as the same way I think of astrophysics. After reading Steven Hawkins book, I still don't really know astrophysics at all, but I have a good enough model of things. And as they say, all models are wrong, some are useful. If I were a lawyer using these things for work, I'd be insane to trust one at this stage, but the reality is I'm not using my digging into things I don't know about for anything load bearing, but even if I were, I'd still use an LLM to get started. Eg the post didn't state how the author learned anything the name for the dropped letter O, but I can describe a thing and have the LLM give me the name of it. there's an emphasis on getting things totally 100% right does erode trust, but you get a sense for what's could be a hallucination and then check background resources if you get enough experience with the tool.