back to article GitHub Copilot is AI pair programming where you, the human, still have to do most of the work

GitHub on Tuesday unveiled a code-completion tool called Copilot that shows promise though still has some way to go to meet its AI pair programming goal. If you're wondering how well it performs, in an FAQ about the service, which is available as a limited "technical preview," GitHub admitted: The code it suggests may not …

  1. IamAProton

    Given the "complexity" (big, fat euphemism) of then code written by some human "developers" I wonder if this autocomplete will make any difference.

    1. Flocke Kroes Silver badge

      I am sure the AI will suggest the most popular patterns. As ignorance outnumbers competence then the sign of difference will depend on the human. The actual use of this software might actually be to identify useful programmers: If they use a large amount of recommended code then they lack the skill to understand what is wrong with it.

  2. amanfromMars 1 Silver badge

    A small but important point to contend is not possible nor highly likely.

    "We use telemetry data, including information about which suggestions users accept or reject, to improve the model," the Microsoft-owned biz said. "We do not reference your private code when generating code for other users."

    Should that be written in the interest of greater accuracy for clearer transparency .... We do not reference Microsoft-owned biz use of your private code when generating code for other users ‽ .

    It is Microsoft after all, which be infamous for its supposed Base DNA Proclivity of Embrace, Extend and Extinguish.

    And if that blanket assumption be totally wrong, I apologise for the popular presumption of their unproven guilt in such matters as matter greatly.

  3. David Harper 1

    The hideous offspring of HAL-9000 and Clippy

    "I see you're trying to write a mission-critical program. Would like help with that?"

  4. Gene Cash Silver badge

    So basically...

    It's automated copy-n-paste from StackOverflow?

  5. elsergiovolador Silver badge

    Stackoverflow 2.0

    People don't even take time to read Stackoverflow solutions, just copy code from the suggestions one by one until something works.

    This could be useful though if you write a test and then the AI will "write" code until the test is passed.

    There is going to be a third excuse why do you watch cat videos, after "it's compiling!" and "it's running tests!" - "the code is writing itself, what am I even doing here!"

  6. Howard Sway Silver badge

    The code it suggests may not always work, or even make sense

    Yet more "software development is just coding" bullshit.

    Why should a programmer want their thought flow interrupted by some bullshit tool offering up unsolicited random suggestions whilst they're busy carefully creating solutions to domain specific problems? A tool like this could never understand the context in which I'm writing a function or class, nor could a bunch of random from hundreds of other developers ever fit in with my own personal design and programming style, developed over decades. I mean, it might compile and work, but who wants a project of any complexity that contains 50 wildly different programming styles?

    Proper coding is design too. Let me know when one of these bots does something actually creative, not just once but consistently.

    1. elsergiovolador Silver badge

      Re: The code it suggests may not always work, or even make sense

      > Why should a programmer want their thought flow interrupted by some ~~bullshit tool~~ other programmer offering up unsolicited random suggestions whilst they're busy carefully creating solutions to domain specific problems?

      FTFY

      And yet many developers love pair programming. I never understood why... to cover each other shortcomings and to try to last until the next payday?

      If manager throws one programmer at the problem and he or she is struggling, then they may not know whether the problem is difficult or programmer is not worth their salt, but if they throw two programmers at it, what are the chances that both are <insert a common name of a waste product expelled by animals>?

      Now it will be "Boss, Co-pilot can't solve it either! We need a quartet!"

    2. Warm Braw Silver badge

      Re: The code it suggests may not always work, or even make sense

      I've always had an issue with the XP approach that you sit in front of a screen and start writing code - specifically a test for some further code you have yet to write - without actually giving much prior thought as to how that code might operate.

      I can see that it might work for routine cases - for example grabbing some stuff from a database at the server side and shipping JSON back to the client or at the client updating the UI from the response - but for any significant problem you need to think in advance about data representations, research algorithms with suitable performance and design interfaces that enhance scalability and maintenance. I still cling to the belief that the process starts in a notebook, not on a screen.

      The "parse_expenses" code is a good example of all the things you need to decide in advance rather than just assume: how might dates be represented; if it's an "expense" what's the significance of +ve and -ve values; will you reimburse based on the exchange rate on the day of the transaction, on the day it was converted (for example paying card bill) or an arbitrary rate? In real life, the "business logic" is unlikely to be something an AI system has seen before and can't help you foresee.

      And in the JSON case, data serialisation is not something programmers should normally have consciously to be dealing with: if it's not built into the language it should be handled by a library or a framework. Always better to reduce the code that needs to be written than automatically generate boilerplate that may or may not correctly infer what it needs to do,

      However, it might be fun to couple two instances of Copilot together, toss in a few lines of code and see what they eventually come up with.

      1. Cederic Silver badge

        Re: The code it suggests may not always work, or even make sense

        I suspect that's because XP never supported sitting in front of a screen and starting to write code without giving prior thought as to how that code might operate.

        Disliking 'big upfront design' doesn't mean doing no design. Those conversations happened over two decades ago, check the c2 wiki for contemporary discussions.

        Pair Programming is advocated within XP but works beyond it too. This isn't pair programming, this is code completion, lacking context and understanding of the problem domain, ignorant of the intended design and in its current form feels a horrible distraction rather than a constructive aid.

        That doesn't mean it can't improve, but it's not pair programming, and it has as little to do with XP as your description of it.

  7. Mike 137 Silver badge

    Quite apart from c*ckups

    This AI may be pretty incompetent, but it's only the peak of an iceberg of useless automation. For example, I've spent literally hours working out how to turn off all the "helpful" automation in Net Beans. Some of it just won't go away, none of it proved useful and some of it is darned stupid (e.g. a popup containing one line of guidance and the statement "Alt-Enter shows hints". Hitting Alt-Enter pops up an identical line of guidance). And let's say nothing about default enforced code formatting to a horribly unreadable standards.

    When I'm programming, I want [a] to be entirely in charge, and [b] to have absolutely no unnecessary distractions. This IDE (as an example) seems to be designed for coders who don't really know much about programming, and I have confirmation of this situation in the number of "<insert language> interview questions" web sites around.

  8. Anonymous Coward
    Devil

    GitHub Copilot doesn’t actually test the code it suggests, so the code may not even compile or run

    The Copilot is a Millennial.

  9. Cuddles Silver badge

    Managing expectations

    "The most enthusiastic person we could find toward Copilot said the tool worked as exactly expected one time in ten. "When it guesses right, it feels like it's reading my mind," they said."

    Based on the descriptions in the article, it sounds like it works exactly as expected at least nine times in ten. What the person quoted appears to mean is that it works as intended one time in ten.

  10. JDX Gold badge

    Easy to be cynical but...

    ... I'll watch with interest how this progresses. Even if it's "only" like a more advanced autocomplete, well A/C saves me a huge about of time and effort on a daily basis.

  11. Michael Wojcik Silver badge

    OpenAI still wasting resources, I see

    We could see it being used to automate easy, repetitive tasks – such as generating boiler-plate code

    That should already be automated, or refactored out of existence. There's no excuse for software developers to be creating boilerplate code. That's either a terrible use of resources, or those developers aren't worth employing.

    and suggesting library or framework functions you're not familiar with to save you the trouble of looking through documentation

    As others have already pointed out, we have StackOverflow copypasta for that. It's an abysmal idea. The return on the cost of consulting documentation isn't primarily in finding an API or getting the details of how to invoke it; it's in learning about the intent behind it and recommendations for its use. Yes, documentation is often poor – but that's important to know, too.

    The majority of effort in programming should be in thinking, then in research. Actually writing code should be down near the bottom. Measure twice, cut once.

    I'm reminded of Djikstra's complaint about APL. Using ML to write code is exactly the wrong approach; it's applying massive resources to a wild misconception of the actual problem.

    1. JDX Gold badge

      Re: OpenAI still wasting resources, I see

      >>There's no excuse for software developers to be creating boilerplate code

      Sorry but this is very naive. Any time you work against someone else's bespoke interface that doesn't happen to use standard web-services, you end up writing this kind of boring stuff. Even if it's only writing a class and mapping property names to JSON so the auto-serializer will do the rest for you.

      You can write a tool to do it, but that often takes longer than the drudge-work.

  12. HAL-9000
    Gimp

    Autocomplete+

    Probably the best one line summary I've seen in a while: Copilot may find other ways to ruin your day.

  13. Anonymous Coward
    Anonymous Coward

    Just what I want when programming

    A random idiot constantly shouting half-arsed suggestions at me.

    A decent, non-AI autocomplete would do me, thanks.

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Other stories you might like