2084: WikiGPT
Imagine a ChatGPT that works like Wikipedia - Community moderated information
The paper I read yesterday on editing the facts contained in a LLM(Large Language Model), got me thinking on how that could be usefully applied. One of the most often complaints about LLM are that they are inaccurate, that they provide incorrect information or hallucinate new information that could never exist. But this isn’t something which always has to be the case. Given that as the paper says, you can edit the facts stored in a LLM, you could take an inaccurate language model, and over time transform it into a more and more accurate language model, by rewriting what it outputs in response to given prompts.
However, doing this directly would take forever, given the sheer amount of facts and knowledge that exists and the necessity of being exhaustive. But this is not a new problem. Wikipedia, one of the most comprehensive websites on the planet, has long had a system whereby it could collect and verify huge amounts of data, using an army of unpaid volunteers. Given that ChatGPT already has millions of users, there must be enough users there willing to put in the time to help ChatGPT, that you could reimplement the system used at wikipedia, and have a hierachy of moderators, along with a talk page, and all the rest, but in this case use it to create a database of facts, which could be loaded into the model. Over time, as users use the model, and improve its store of facts, the model could become more and more accurate, and in fact would probably be greatly expanded. It could be something like where you could flag inaccruate information in a response, and submit it along with a source, or it could be a more complicated system, but the most important part is that it is already well within the realm of possibility. In 2084, you could have absurdly advanced LLM on a variety of subjects, each of which would be constantly checked to make sure that all the information it returns is as accurate as possible.
