4 ms·

> esp. those in the tech community It’s a vocal minority of the tech community that shares that opinion. The silent majority knows it’s just fud marketing. Per

• jiggawatts 3y ago
  Few people are "blown away" by the absolute skill level of ChatGPT. It's more that there exists an AI that can do this at all. This just wasn't a thing three years ago, and the first truly capable version (GPT 4) only become publicly accessible this year. Some things I've seen GPT 4 do is "as good as human", often better than average. For example, spotting bugs or reverse-engineering code and then writing doc comments for it. The real power is not in having an AI as smart as you, it's having something "good enough" that can be deployed cheaply at scale to do rote, menial tasks. Review code, check for vulnerabilities, fix spelling mistakes in comments, etc...
  • webmaven 3y ago
    "The marvel is not that the bear dances well, but that the bear dances at all." — Russian proverb
  • chmod775 3y ago
    You want to trust that thing to review your code? It will often give you eloquent explanations, trying to convince you that code has a certain bug, while completely missing the actual bug. You'll be gaslit into chasing down dead ends, distracted from actual issues. Example: "In JS, create a Promise that is resolved with some emitted value from an EventEmitter, or rejected if an 'error' event is emitted first. Write that concisely, using '.once' and only removing the other event respectively." Fairly everyday thing in node. Every competent programmer would grasp the intent here pretty quick and write correct code, but somehow it's generally beyond ChatGPT. Example ChatGPT output (comments mine): function eventToPromise(emitter, event) { return new Promise((resolve, reject) => { emitter.once('error', reject); // memory leak: not removing the other listener emitter.once(event, (data) => { emitter.removeListener('error', reject); resolve(data); }); }); } Come on now... I just start the whole prompt over to let it try again: function promiseFromEvent(emitter, eventName) { return new Promise((resolve, reject) => { const resolveHandler = (data) => { emitter.removeListener(eventName, errorHandler); // memory leak: should be removeListener('error', errorHandler) resolve(data); }; const errorHandler = (error) => { emitter.removeListener(eventName, resolveHandler); reject(error); }; emitter.once(eventName, resolveHandler); emitter.once('error', errorHandler); }); } Now I prompt "You have a bug in there, fix it." ChatGPT: "Apologies for the oversight. You are right; there is a bug in the previous code. The issue is that if the 'error' event is emitted first, both the 'error' and 'data' event listeners should be removed, and the Promise should be rejected." Nope. The error event listener is actually the only listener working correctly... It proceeds to introduce a cleanup function that removes both listeners (despite keeping '.once'), fixing the bug by accident, not intent. Asking it to change it so that the code adheres to the original prompts starts a downward spiral.
    • jiggawatts 3y ago
      90-95% of programmers I've worked with can't "grok" asynchronous or multi-threaded programming, and when forced to write async code it is littered with bugs. ChatGPT doesn't need to find 100% of bugs in 100% of code out there. Think of it as a "better linter". It's cheap, and finding just 50% of bugs would be an enormous step up for quality in most code bases.
      • donmcronald 3y ago
        > Think of it as a "better linter". No way. A linter is deterministic. ChatGPT is all over the place and, for me, it’s wrong almost every time I ask it anything. I wouldn’t trust it to tell me the ingredients in a pepperoni pizza. I’m definitely not letting it give me programming advice.
        • jiggawatts 3y ago
          Ironically, I had a lot of trouble with a particular recipe (pan-fried gyoza with a crispy bottom), and it was only GPT 4 that gave me a working recipe! The lack of determinism can be considered a type of strength. Run it multiple times! It might find different bugs each time. Humans are the same, by the way. If you show a random set of programmers random snippets of code, you'll get a non-deterministic result. They won't all find all of the bugs.
          • donmcronald 3y ago
            > The lack of determinism can be considered a type of strength. Run it multiple times! It might find different bugs each time. I’ve only tried it with code a little bit, but what I find is that it gives me hallucinations that I need to spend time figuring out. I don’t know what it’s saying, because it’s gibberish, and then I have to spend my time figuring out it’s not accurate. I don’t want to run on that treadmill. I’m guessing software will continue the trend of getting less reliable as more people are willing to generate it via an AI.
            • 3y ago
  • moonchrome 3y ago
    By the time you run enough iteration loops to chunk things up and summarize + context size it's neither cheap nor fast when you account for the effort it takes to get it there. I could see tech support eventually being profitable but in my experience even GPT4 is just so randomly unreliable - it's not there yet.
• ksec 3y ago
  >The silent majority I only wish this is true. Because so far "The silent majority" are no where to be seen in many other subjects.