Tell HN: I asked ChatGPT to build a login form in React and it gave me the code

It's annoying blog spam that will hijack your attempt to scroll back with a modal asking you to enter an e-mail and subscribe, but it seems to be roughly the same thing, except the entire article appeared at once and I didn't need to interact with it and wait five minutes for the full answer.

Here's how it went: https://i.imgur.com/PS6oPsq.png

A pretty incredible developer experience, if you ask me!

Well, mostly. I tried the prompt in the OP into text-davinci-003 and got this:

    const handleLogin = (e) => {
      e.preventDefault();
      // check if the email and password are valid
      // perform login
      console.log(email, password);
    };

1. those that can specify concise requirements for the AI

2. those that can check correctness of the code the AI generated.

I did ask it something more basic: "how to use fetch to download and parse a csv", and the code was helpful, although not much different from what I could find with google/stackoverflow. However, this next part was what really impressed me. The results used Papa for parsing, so I then asked it to do it with csv-parse instead, and it worked! That seems like the advantage over something like Copilot--it has contextual history of the question which can be used conversationally to tweak the results.

Prompts:

- Write a simple CRDT implementation in Swift

- How can I make a GraphQL server using Rust using the async-graphql and actix-web libraries?

- Write code to create a Dropbox clone in Rust

- How do I make an HTTP request in Javascript?

However, just like with Copilot, it can introduce subtle bugs. For example, in the Dropbox example, notice how it uses synchronous code in an async function. Also, longer prompts will reach the token limit and get cut off.

Interestingly, this must be what Replit is also using for their AI codegen service too (https://replit.com/site/ghostwriter) since I believe they have some connection to OpenAI, as they got early access to OpenAI's service last year with their previous Codex product (https://blog.replit.com/codex).

One easy way of catching out is to ask "write X in language Y", where X isn't something most people would write in Y -- it usually got it wrong for me.

Here's my convo with ChatGPT: https://imgur.com/a/eIjvh8C

If the text and code are just random stuff making no sense then I wonder how different it is from the Copilot results (besides the convincing style of answers).

For a few of them it made a small error of one line that I had to fix. But also in a few cases I was able to just reply that there was a bug and what it was doing incorrectly and to fix it, and it worked.

It gave me code that was very specific such as using the methods as I named and described them in the request for the class.

The big question is: whose code is this? Is it, verbatim, someone else's copyrighted code (assuming a snippet like that is novel enough to be copyrightable)? Or is it derived from someone else's code? And if it's derived, is it a derivative work in the legal sense, or is it no more derivative than if a human had read a tutorial or some documentation and then later written code like it?

From a practical perspective I do think a lot of programmers are doing copy-paste in their day jobs, which means they should be reusing existing code (I would imagine there are numerous reusable customizable login forms).

---

For example I asked it this:

Can you write a javascript function that takes a callback and returns the value of that callback, but returns a cached value if that function has been called less than 5 seconds ago?

And it produces code that looks right but doesn't work (doesn't handle arguments, failing its own example implementation). However, upon reminding it to support arguments it does work. I guess this is actually potentially useful.

Maybe it can substitute StackOverflow, we will get better and faster code / explanation. But in the next 30-50 years, who knows. I guess it can write a lot of code for us, maybe we will be obsolete in that period of time, but as we know, prediction of the future often times goes really really wrong.