Using Generative AI for programming

Using Generative AI For Programming

The rise of generative AI

  • Generative AI is the new hotness in the whole IT field
  • Generative AI models
    • …learn the patterns and structure of their input training data
    • …generate new data that has similar characteristics
    • a.k.a. write a prompt, and you get text/code/images/videos/music/etc
  • Large Language Model (LLM) is a form of generative AI
    • Used often in a chat-like interface
    • OpenAI’s GPT series is a LLM, a technology that fuels the popular ChatGPT chatbot

GitHub Copilot

  • GitHub’s paid code auto-completion service that is advertised as AI pair programming
  • What is it really?
    • Autocomplete on steroids, $10/month
    • See Copilot here in action
    • It looks like magic, right?
    • This could be useful for generating basic boilerplate code
      • But wait, there’s var variables there…
      • Old JS! Shouldn’t be used nowadays!
      • You might know this, but not everyone knows that

How does generative AI work?

  • All AI models need data
  • AI does not know when it’s wrong, when it’s using bad practices
    • TBH, it does not know anything
      • It just returns an answer that best matches your prompt
      • It might be exactly what you need, but it also might not
  • The future of AI is worse: At some point, AI will be taught on AI-generated code
    • The quality of the result worsens!

Is ChatGPT any good?

  • Okay, we know the risks. What’s ChatGPT good for?
  • Let’s ask ChatGPT!
  • Summary:
    • Makes mistakes
    • Doesn’t understand context
    • …And I somewhat disagree the pros 3-5

bg right width: 87%

Using ChatGPT professionally

  • Do note that GPT-3 and GPT-4 (the paid version, $20/month) are not equal in terms of quality of results
  • I asked two senior developers who’ve used ChatGPT professionally for its pros and cons
  • Some comments I got:
    • “I don’t usually use it for writing code, but for other stuff”
    • “It makes some mindless tasks way faster”
    • “Not always right, but surprisingly often!”
    • “Works for some languages, but not all”

Example use cases (Dev 1)

1) Finding causes for error messages * ChatGPT can the cause for a given error message faster than a developer * You give code & error message -> ChatGPT gives fixed code as a result * Gotcha: This works when you are working on something that someone else has worked on before

2) Explaining architectural solutions * Examples: * “How to set up a service in AWS for x number of people” * “How to set up an event based service”

Example use cases (Dev 1), cont.

3) Reading documentation for you

  • Make a summary of a long documentation! 4) Writing regexes 5) Writing helper functions
  • “Make sure this string is 4 characters long”
  • “Give me input validation for x”
  • Always test the results with care

Example use cases (Dev 2)

1) Give high-level specs for a program * “I should do software in framework X that should do Y” * ChatGPT gives a basis for the software * This is a good starting point for writing your software * …but rarely even a single line of the basis are left in the final software! 2) Help you understand code you can’t fathom at all * “Convert this to Python!” 3) When you have to write something you’ve never written before

  • But don’t trust it on this
  • If the programming language is strict about its syntax, ChatGPT results are generally better
    • Rust is good, C++ bad
    • Also: Does not work with C++ well because there is so much badly written or obsolete C++ on the Internet!

Let’s also remember these words from Linus

Generative AI is here to stay

  • I don’t believe for a second generative AI will make programmers obsolete
  • It won’t go away, though, and it will change the nature of some programming
  • It’s a new tool, and in a field where we get new tools every year, it’s good idea to keep your mind open for new tools
  • It’s not a silver bullet, however
    • It can make you bypass learning altogether
      • And you NEED to understand code in order to use AI tools effectively
    • It can make you write worse code
      • …but it can also make you a faster programmer.
      • …and a faster programmer is a better programmer!

AI is a good servant, but a bad master

Reading & Watching