So much jumping to defend llms as the future. I'd like to point that llms hallucinate, could be injected, and often lack context which well structured metadata can provide. At least, I don't want for an llm to hollucinate the author's picture and bio based on hints in the article, thank you very much.
I don't think that one is necessarily better than the other, but imagining that llms are a silver bullet when another trending story in the front pages is about prompt injection used against the slack ai bots sounds a bit over optimistic.
Prompt injection is highly relevant because you end up achieving the same as the publisher choosing the metadata, but on a much higher price for the user. Price which needs to be paid by each user separately instead of using one already generated.
LLMs are much better when the user adapts the categories to their needs or crunches the text to pull only the info relevant to them. Communicating those categories and the cutoff criteria would be an issue in some contexts, but still better if communication is not the goal. Domain knowledge is also important, because nitch topics are not represented in the llm datasets and their abilities fail in such scenarios.
As I said above, one is not necessarily better than the other and it depends on the use cases.
> Prompt injection is highly relevant because you end up achieving the same as the publisher choosing the metadata, but on a much higher price for the user.
How does price affect the relevance of prompt injection? That doesn't make sense.
My question is: how price does not matter? If you are given the choice to pay either a dollar or a million dollars for the same good from an untrustworthy merchant, why would you pay the million? And the difference between parsing a json and sending a few megabytes of a webpage to chatgpt is the same if not bigger. For a dishonest seo engineer it does not matter if they will post boasting metadata or a prompt convincing chatgpt in the same. The difference is for the user.
I don't mind the delusions of most people, but the idea that llms will deal with spam if you throw a million times more electricity against it is what makes the planet burning.
I don't think that one is necessarily better than the other, but imagining that llms are a silver bullet when another trending story in the front pages is about prompt injection used against the slack ai bots sounds a bit over optimistic.