3 ms·

Flying with GitHub Copilot

1 points by dtagames 4y ago

verdverm 4y ago
The removing duplicates section is the kind of thing that has me worried about technologies like this
- dtagames 4y ago
  What aspect of that is worrisome?
  - verdverm 4y ago
    On the surface, the algo complexity Below the surface, that it's called clever by the author Deeper, the issue with humans deferring to AI without thinking for themselves. This is a problem with AI generally. See self-driving cars and doctors learning to appease the image scanning algos
    - dtagames 4y ago
      I'm the author and I specifically pointed out that using it to learn and making sure you understand are the best use cases.
      - verdverm 4y ago
        The section I referred to is an example of where Copilot goes wrong. Are you early enough in your programming journey that algorithmic complexity hasn't come up yet?
        dtagames 4y ago
        More than 25 years in, so I would say no. What problems do you see with the suggestion, or what could be done to improve it?
        verdverm 4y ago
        Ah, I have misinterpreted "using it to learn" from a few comments ago I would use the "remove duplicates" section to illustrate that Copilot doesn't always produce quality code, and that you should always be mindful and really evaluate what it is suggesting. https://www.rosettacode.org/wiki/Remove_duplicate_elements https://www.rosettacode.org/wiki/Remove_duplicate_elements would be a useful link to include. The Copilot example in the post is in the N^2 camp I'd personally go further and describe how the issues we see with NLP models encoding our racism and biases is similar to how OpenAI Codex picks up bad coding examples practices. Perhaps that deserves a post unto itself? Or a general criticisms of Copilot post? (because there are quite a few now :)
        dtagames 4y ago
        I disagree that the remove duplicates suggestion is bad. As I mentioned, I think it's a great suggestion. Can you write a better version and post it here in JS? I'd love to see why you're unhappy with it. On the racism issue, I understand that is an issue with facial recognition and such, but its hard to imagine any racial bias is going to creep into anything I write. Would Copilot suggest that all the 'customer' objects in my example have a race? What are you aiming for here? I see that you work at a company that makes a tool like Copilot. Is the complaint that you feel your company's tool is better?
        verdverm 4y ago
        Suggesting bad runtime algorithms is the problem. You should try writing it so you understand algo complexity. If you do not understand Big O concepts after 25 years, I'm not going to try explaining it to you. Racism & biases in NLP is only an analogy, all ML models have issues with capturing undesirable features of the training data. My complaint is about you suggesting bad ideas are good ones. The `hof` tool is more like Helm than Copilot. My opinion is that deterministic algorithms are often better when the solution space is represented by graph-like structures. ML is not advanced or capable enough yet for problems that require exactness across time & space. I am interested in applying Copilot like models to the design space in `hof` (config/data rather than code)

1 points by dtagames 4y ago

verdverm 4y ago

The removing duplicates section is the kind of thing that has me worried about technologies like this

dtagames 4y ago
What aspect of that is worrisome?
- verdverm 4y ago
  On the surface, the algo complexity Below the surface, that it's called clever by the author Deeper, the issue with humans deferring to AI without thinking for themselves. This is a problem with AI generally. See self-driving cars and doctors learning to appease the image scanning algos
  - dtagames 4y ago
    I'm the author and I specifically pointed out that using it to learn and making sure you understand are the best use cases.
    - verdverm 4y ago
      The section I referred to is an example of where Copilot goes wrong. Are you early enough in your programming journey that algorithmic complexity hasn't come up yet?
      - dtagames 4y ago
        More than 25 years in, so I would say no. What problems do you see with the suggestion, or what could be done to improve it?
        verdverm 4y ago
        Ah, I have misinterpreted "using it to learn" from a few comments ago I would use the "remove duplicates" section to illustrate that Copilot doesn't always produce quality code, and that you should always be mindful and really evaluate what it is suggesting. https://www.rosettacode.org/wiki/Remove_duplicate_elements https://www.rosettacode.org/wiki/Remove_duplicate_elements would be a useful link to include. The Copilot example in the post is in the N^2 camp I'd personally go further and describe how the issues we see with NLP models encoding our racism and biases is similar to how OpenAI Codex picks up bad coding examples practices. Perhaps that deserves a post unto itself? Or a general criticisms of Copilot post? (because there are quite a few now :)
        dtagames 4y ago
        I disagree that the remove duplicates suggestion is bad. As I mentioned, I think it's a great suggestion. Can you write a better version and post it here in JS? I'd love to see why you're unhappy with it. On the racism issue, I understand that is an issue with facial recognition and such, but its hard to imagine any racial bias is going to creep into anything I write. Would Copilot suggest that all the 'customer' objects in my example have a race? What are you aiming for here? I see that you work at a company that makes a tool like Copilot. Is the complaint that you feel your company's tool is better?
        verdverm 4y ago
        Suggesting bad runtime algorithms is the problem. You should try writing it so you understand algo complexity. If you do not understand Big O concepts after 25 years, I'm not going to try explaining it to you. Racism & biases in NLP is only an analogy, all ML models have issues with capturing undesirable features of the training data. My complaint is about you suggesting bad ideas are good ones. The `hof` tool is more like Helm than Copilot. My opinion is that deterministic algorithms are often better when the solution space is represented by graph-like structures. ML is not advanced or capable enough yet for problems that require exactness across time & space. I am interested in applying Copilot like models to the design space in `hof` (config/data rather than code)