LLM assisted development has made me at least twice as efficient. I'm using GitHub copilot, CodeGPT (Visual Studio Code extension with ChatGPT4), and Llama indexes trained with the latest documentation of the tools I'm using (mainly useful for dbt where a lot of new features have been released in the last 18 months).
It's good at syntax and established patterns. Which means I can spend more time on interesting and creative problems like finding new novel and efficient patterns. It's not going to replace my job any time soon but it is helping me solve more problems.
What other LLM tools should I investigate? Are you using LLM assisted development? If so in what fields and what tools are you using?
@greg I use copilot, but I refuse to use CodeGPT since that would essentially upload my clients' confidential code to the ether.
Copilot's not bad, but it's only right about half the time.
@LouisIngenthron It depends how you use it. You can still get value from LLM assisted development with ChatGPT without uploading your clients code
@greg The AI runs on the provider's server, does it not? How could it possibly provide suggestions without it uploading the code to the provider's server for analysis?
@LouisIngenthron Prompt it with something like "write a python function that takes a list of lowercase surnames and returns them capitalized? Don't forget to appropriately capitalized names like McDonald etc". In this example it extrapolated that I would also want names like O'Connor appropriately capitalized. I can do this myself with Google and Stackoverflow but it's much quicker with LLM tools. No need to upload private code to get value from them
@greg Certainly not, because I only advertise my services for languages I already have experience in.
If you don't know the language you're programming in and you're just prompting an AI and publishing, then why couldn't your client cut you out and just prompt the AI to do what they want themselves? If you, by your own admission, don't know the language, then you're not really adding much to that equation, are you?
@LouisIngenthron my value as a developer isn't in my ability to memorize syntax, it's my ability to generate efficient code. If a client could solve their problems by blindly prompting ChatGPT then I would encourage them to do that. It would be a waste of my skillset. The value I add is in knowing if the provided solutions are useful, knowing what is possible, etc
@greg How can you know if the provided solutions are useful if you don't know how the language works? That sounds like a recipe for, just as an example, C++ code riddled with memory leaks.
Knowing a language is about more than just syntax; each language has its own paradigms and pitfalls to look out for.
Delivering code written by an AI in a language you don't know to a client for money feels deeply unprofessional to me.
@LouisIngenthron
I assume you use Google, Stackoverflow, etc when you develop. I also assume you have the experience to know if a solution you find is appropriate for your use case. How is utilizing LLMs to help with syntax or patterns any different from using existing tools. I don't understand your hesitation to use a new tool that can help you be more efficient.
@greg In languages I know, I have the requisite experience to make those determinations, yes.
@LouisIngenthron so why won't you add conversational LLMs to your toolkit? Honestly is smells like pride more than a reasonable argument against the tool.
@greg Because I work in large codebases and it would violate my clients' privacy to allow CodeGPT to upload those to OpenAI.
Also it would be deeply unprofessional to pass off low-quality AI-driven code as my own.
I'll use AI to speed up my work. Not to create it whole cloth.
@LouisIngenthron that's simply not true. Surely you come across instances where you need to implement some functionality in a new to you language. In those cases, giving an LLM the concept and letting it deal with the syntax is time saving. I appreciate you can do some of that in copilot but why not use all the tools at your disposal? Conversational prompting is far more effective.