4 ms·

I am waiting for the day that GPT can show it's work so to speak. i.e. "Debug last results" and get a list of all the trained input data including any URL's, b

I am waiting for the day that GPT can show it's work so to speak. i.e. "Debug last results" and get a list of all the trained input data including any URL's, books, text input and who performed the input, where the data came from. I am told this will never happen but I think courts will disagree once enough big mistakes and/or potential malfeasance or fraud occurs and people try to blame GPT. Or when GPT repeats the unsavory behavior we have all witnessed with social media algorithms leading to highly divisive outcomes.

softfalcon 4y ago
I think this is a good idea. When we disagree about things in conversation, we as humans ask "where did you hear that?" and from there, we dig into the source of the thought process. Unfortunately, a lot of what folks are saying to you is true. It's very hard for GPT to cite its results because the way it is trained it is by design made to garble the inputs down into a simplified concept of guessing the next words given some input. It essentially just has a procedural generator for words that is seeded by your prompt. If you give "alfalfa" it will head off in a direction towards "farm", "hay", and "grazing" along with other connective words to form sentences. Because its concept of data is all around just word spaces, it can't really go "oh, I read about alfalfa on a Wikipedia article with x sources". It just knows "alfalfa is like the word grazing". I am simplifying to make a point, but this is in essence how these algorithms work, directions of traversal, guided by probability, towards word clouds floating in a grouped space. This is sort of changing though. Bing and Google (as well as many other researchers) are using specialized databases to provide further context that is fed into your prompts that come from real search results. Theoretically, they could get this tuned enough that GPT and other LLM can have the right data to provide a connection to cited facts alongside the hallucinated glue language. I feel like what you're asking for is valuable, but might take a bit before we really get it relatively accurate.

softfalcon 4y ago

I think this is a good idea. When we disagree about things in conversation, we as humans ask "where did you hear that?" and from there, we dig into the source of the thought process. Unfortunately, a lot of what folks are saying to you is true. It's very hard for GPT to cite its results because the way it is trained it is by design made to garble the inputs down into a simplified concept of guessing the next words given some input. It essentially just has a procedural generator for words that is seeded by your prompt. If you give "alfalfa" it will head off in a direction towards "farm", "hay", and "grazing" along with other connective words to form sentences. Because its concept of data is all around just word spaces, it can't really go "oh, I read about alfalfa on a Wikipedia article with x sources". It just knows "alfalfa is like the word grazing". I am simplifying to make a point, but this is in essence how these algorithms work, directions of traversal, guided by probability, towards word clouds floating in a grouped space. This is sort of changing though. Bing and Google (as well as many other researchers) are using specialized databases to provide further context that is fed into your prompts that come from real search results. Theoretically, they could get this tuned enough that GPT and other LLM can have the right data to provide a connection to cited facts alongside the hallucinated glue language. I feel like what you're asking for is valuable, but might take a bit before we really get it relatively accurate.