7 ms·

I think it could be possible to feed the LLM with lots of data about your product, potentially even give it source code with instructions not to share it. Then

I think it could be possible to feed the LLM with lots of data about your product, potentially even give it source code with instructions not to share it. Then have it act as a highly knowledgeable support rep for customers, and also be able to file detailed issues for the dev team when it identifies that a customer finds a bug. The risk of hallucinations is still there, but it could probably be mitigated by feeding the LLM's output into a fact-checker model. I think the bigger risk is leaking internal data through prompt injections

mooreds 2y ago
We engaged a startup (kapa.ai; they're great) to train an LLM on: * our public facing docs * our public repos * our youtube channel * our forum and spit out answers. It's not right all the time, but it is most of the time and, crucially, it shares not just an answer but a link to the docs that led it to that answer. That makes it easier to fact check the LLM. We've found it very useful for scaling support and making that corpus of knowledge available to folks both outside and inside the company.
autoexec 2y ago
> I think it could be possible to feed the LLM with lots of data about your product, potentially even give it source code with instructions not to share it. I wouldn't trust the LLM to keep your secrets. There are countless examples of people convincing LLMs to say things they shouldn't. On the other hand, I wouldn't trust an LLM to spit out your source code without hallucinating a bunch of fake functions either, so maybe it balances out.