> There is nothing wrong with the LLMs, you just have to double-check everything.
That does not seem very helpful. I don't spend a lot of time verifying each and every X509 cert my browser uses, because I know other people have spent a lot of time doing that already.
The fact that hallucinates doesn’t make it useless for everything, but it does limit its scope. Respectfully, I think you haven’t applied it to the right problems if this is your perspective.
In some ways, its like saying the internet is useless because we already have the library and “anyone can just post anything on the internet”. The counter to this could be that an experienced user can sift through bullshit found on websites.
A argument can be made for LLMs; as such, they are a learnable tool. Sure it wont write valid moon lander code, but it can teach you how to get up and running with a new library.
Ask not what the LLM can do for you. Ask what you can do in order to prompt the LLM better so it can finally produce the correct result. The more that happens, the more it can learn and we can all win.
Think of it like voluntarily contributing your improvements to an open source library that we can all use. Except where the library is actually closed source, and controlled by a for-profit corporation.
Train the LLM by feeding it all your data. Allow it to get better. We all win from it. Maybe not today, but one day. It may take your job but it will free you up to do other things and you will thank your LLM overlords hehe
Wait, how does rewriting a prompt until it gives you the output you expect help the LLM learn? Are you suggesting better prompting gets fed back into the training process in some helpful way? This feels confused.
You think OpenAI isn’t using your prompts and results to train better models? Think of yourself as a large RLHF experiment LOL
Kinda like Netflix did with people watching movies 10 years ago. The data’s there, and abundant. People are massaging their chatbot to get better results. You can measure when people are satisfied. So… obviously…
If an official comes to my door with an identity card I can presumably verify who the person is (although often the advice is to phone the organisation and check if unsure) but I don’t necessarily believe everything they tell me
That does not seem very helpful. I don't spend a lot of time verifying each and every X509 cert my browser uses, because I know other people have spent a lot of time doing that already.