I'm glad they go in deep on the data feed back on this.
Kurzgesagt made a great video on how this stuff gets generated, an article uses this, data sets get trained on this, and then future outputs will use the article as citation.
This is the paradox to be addressed. In trying to make LLM's more accurate, we try to pin it to citations, but then that means having to trust that the citations are accurate. But having the citations generated by an LLM means it basically becomes impossible to confirm the accuracy.
As it has been said anything after 2022 cannot be reliably trusted. You can do you due diligence but that is a lot of work if you can do it.
Really highlights the idea that a lie will travel around the world before the truth puts its shoes on.
ColdStream•42m ago
Kurzgesagt made a great video on how this stuff gets generated, an article uses this, data sets get trained on this, and then future outputs will use the article as citation.
https://www.youtube.com/watch?v=_zfN9wnPvU0
This is the paradox to be addressed. In trying to make LLM's more accurate, we try to pin it to citations, but then that means having to trust that the citations are accurate. But having the citations generated by an LLM means it basically becomes impossible to confirm the accuracy.
As it has been said anything after 2022 cannot be reliably trusted. You can do you due diligence but that is a lot of work if you can do it.
Really highlights the idea that a lie will travel around the world before the truth puts its shoes on.