Flying with GitHub Copilot | Hacker News Reader

Flying with GitHub Copilot

medium.com

1 pointdtagames11 comments

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

What aspect of that is worrisome?

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

I'm the author and I specifically pointed out that using it to learn and making sure you understand are the best use cases.

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?

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?

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 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 :)

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?

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)

I still am not seeing any reason you don't like the algos it suggested, nor any algo from you that's any better. All heat, no light. So I can't take your complaint seriously. Sorry. If you want to suggest an actual algorithm that's better (and not based on ad hominem), I'd love to hear it! We've had several back-and-forth discussions now where I've written germane comments about the actual output and you have... not.

I've already provided a link in a previous comment which has several algos and discussed the tradeoffs. I've also outlined the issue with the one you presented in the post. Perhaps reviewing the thread would be a useful activity?

https://www.rosettacode.org/wiki/Remove_duplicate_elements

Reply on news.ycombinator.com