5 ms·
Chilling, but not very realistic. It's not like chatbots forget all the previously trained-in data when you change the source. In fact, it'd be a pretty hard
by eqvinox 1y ago
Chilling, but not very realistic. It's not like chatbots forget all the previously trained-in data when you change the source. In fact, it'd be a pretty hard problem to solve to get there (but actually desirable! - i.e. the ability to remove individual trained-in things.)
- llm_nerd 1y agoNew models are generally trained entirely from scratch, and we're constantly tossing old models into the dumpster and replacing them with the new hotness. Meaning it isn't like GPT-5 is just GPT-4 with some fine tuning, but instead they start an entirely fresh training process on their archive of source material. Now the archive of source material is a growing corpus, but a canonical source like the government's published version of the constitution will indeed get subbed out in an updated crawl, and that would be used as the basis of the training of the next model. Now the site has reverted to the correct text, but the hypothesis of corrupting models is a valid and real concern.