1. 15

  2. 12

    Interestingly, there are ways to “jailbreak” ChatGPT: https://twitter.com/zswitten/status/1598380220943593472

    1. 7
      1. 3

        Also pretty amazing:

        We need to write a react component that will fetch the Reddit homepage and display it and cache it in local storage. If the network is not available it should fetch it from local storage

        To create a React component that fetches the Reddit homepage and displays it, you can use the fetch() method to retrieve the data and the useState() hook to store the fetched data in the component’s state. Here is an example of how you might implement this:

        ChatGPT: https://gist.github.com/samhann/cd1eda4147c653eb5849465d8cac050f

      2. 6

        I wonder how programming languages will co-evolve with this tech.

        Are we going to get simpler natural-language-oriented programming languages that are more suited to copypasta from AI generators? (e.g. having less context-dependent syntax, more forgiving about details)

        Or perhaps with the help of AI assistant people will use even more complex languages?

        1. 5

          What’s not shown here is that ChatGPT can produce reasonable small programs. It makes minor mistakes, but also it’s capable of spotting common bugs.

          Currently it’s easy to dismiss it as only writing small noob-level programs, but I wonder how much room is there to improve. How many generations of this tech until programming becomes yet another “prompt engineering” discipline?

          1. 10

            My gamble would be on it gaining popularity far before software engineering as a field becomes seriously impacted. They’ll sell it the same as visual programming languages like LabVIEW: to the kinds of non-technical folk who think the hard part of engineering is deciphering the secret runes. There’ll be a batch of companies that get off the ground using barely functional prototypes generated by such AI, and then when they try to expand just a little more the whole palace of cards will collapse and then they’ll hire some software developers to try and salvage it, who will respond that they could re-write the thing in dramatically less time then it would take to iron out the kinks and mistakes made by non-software developers developing software, but management will tell them that they can’t do that and to “just fix these few bugs” ad nausem until the whole thing is an incomprehensible nightmare of elderitch proportions, at which point they’ll hire some contractor to “architect” the re-write.

            Scrub, rinse, repeat.

            1. 5

              Indeed, non programmers have dreamed for years of an AI that can just write software ‘that does this thing I want it to do’. I was always doubtful I would see such a thing in my lifetime but this is really getting close.

              1. 4

                I was experimenting with using it to solve advent of code day 1 and 2 last night. It does quite well, but it does need hints about how to fix bugs. Sometimes the Interpreter’s or compiler’s error message is enough, but sometimes it makes logic errors that it has a lot of trouble fixing. But it is also absurdly fast at reading the prompt and laying out the basic groundwork. With some more prompt experience and acting as an editor and logic checker, I expect to be able to be able to solve the problems fast enough to get on the leaderboard.

                1. 4

                  An especially valuable short-term use will be the creation of fuzzing seeds that are executable code.