2 ms·
> Moving forward, every app we interact with will have a conversational interface, … I don’t think we have a good idea yet of how to wire up LLMs with software
by layer8 3y ago
> Moving forward, every app we interact with will have a conversational interface, …
I don’t think we have a good idea yet of how to wire up LLMs with software APIs in a safe way. This includes the LLM properly understanding the semantics and context of the software it interfaces with, in relation to the user’s intents.
> …and we will no longer be limited by the bandwidth of how fast our fingers can tap on keys or screens.
A physical keyboard actually tends to have higher bandwidth than speech. Well, maybe not for the mentioned “midwits”, in the author’s opinion.
- htag 3y agoI have a few thoughts on this. * We have to use auth that belong to the user and not the LLM, or we have to use a combination of LLM and user keys. The LLM cannot have global read/write access. * We could give each LLM limited access to APIs but have multiple LLMs. In this way we can control user access to API endpoints based on which LLM they have access to. * Any context building information we feed the LLM can probably be leaked so we shouldn't feed it any sensitive prompts. In fact we should assume any information we give the LLM can be leaked, so we should treat feeding an LLM data as the same security risk as feeding the user data.