Hacker News Reader

Comment by geonnave | Hacker News Reader

My experience is completely different. I have successfully used GPT-4 to:

- write a contract for the sale of my motorcycle: put all details, names and numbers with labels on a spreadsheet, paste on the chat and ask for a contract, then edit.

- learn french: I told gpt "when I write wrong stuff in french, always let me know and teach me the correct ways". Then, after a few weeks I asked for a .csv with the stuff that he corrected so I could import into Anki, which actually worked.

- coding on a daily basis: I am learning Rust on my new job, so I ask it things all the time, it helps me a lot.

View on HN

Your experience is not different from OPs. You're just ok with the mistakes or are unaware of them because you don't know how to judge them.

I both use ChatGPT to boost productivity but also see the amount of mistakes it makes and will keep making and am surprised at the extreme denial of anyone who tries to shut down criticism of the wrong type of hype (the one that sells something that is not there)

Oh but it is: I find it useful. For example, I would not pay a lawyer for that contract, so having it draft me a mediocre contract is still better than having no contract.

A mediocre contract can be worse than having no contract, or as bad as having no contract if it isn't actually legally binding. Yikes.

It can be much much worse if you base key decisions on the belief that you have a contract and later find out that you didn't.

Thanks for such a clear cut example for my points.

Has a lawyer looked over your contract and confirmed that it is actually legally valid? Can you name said lawyer and the firm they work at?

Would you be willing to publish the contract with sensitive information redacted?

All these bold claims are just claims until people come up with some substance. Talk is cheap and confirmation bias happens all the time.

All you need for a bill of sale is a simple sentence saying on x date I xxx, sale vehicle xxx, with vin number xxx to xxx person. Then write the driver's license of both parties and sign. A lawyer is extreme overkill for such a simple transaction.

By the same logic, a hallucinating LLM is also overkill versus just doing the simple task yourself and not needlessly adding risk to it.

The point still remains: let's see what the LLM delivered that the user actually used. Either it's legally binding and an appropriate use, or it's not fit-for-purpose.

Equally, why not an interactive form using conditional logic? No hallucination possible. Much more simple and reliable.

> All you need for a bill of sale is a simple sentence saying on x date I xxx, sale vehicle xxx, with vin number xxx to xxx person.

If you know this, you don't need GPT.

If you don't know this, you don't have a way to assess GPT's attempts at a contract. A bill of sale is indeed simple, but there's a lot of more subtle legal issues someone might run into in life.

Good to know it's able to do some useful stuff. In my case I mostly ask Python related questions, because it's the one I know better so I can check if the answer is right or wrong. I will try with different languages, but I will be less capable of knowing if I got a good answer or not. It may take more time, but I find the combination of Google + Stack Overflow more accurate than asking ChatGPT

Reply on news.ycombinator.com