2 ms·
I suppose that's a relatively easy thing to fix, technically. It proves, however, that th underlying LLM is trained on copyrighted data. I'm not sure the probl
by rich_sasha 3y ago
I suppose that's a relatively easy thing to fix, technically. It proves, however, that th underlying LLM is trained on copyrighted data.
I'm not sure the problem goes away simply if the LLM in question (or any other one) gets some "no verbose regurgitation" filter.
- exitb 3y agoIn that case, the language model calls a search function and just repeats the result out its conversation context, not its training data. With that in mind it's not clear why it's ok for Bing itself to quote the source, but it stops being ok, when a chatbot does it.
- kolinko 3y agoThe example from the article doesn't show that LLM is trained on copyrighted data - it's just Bing fetching the source article, providing it to GPT, and GPT rephrasing the article. An agent trained on entirely copyright-free data would provide exactly the same output.