If you do, what workflows have worked best for you? What models do you trust? Do you believe that it may be a problem offloading thought you would do yourself otherwise to a model?
I use these tools in projects where there is no sensitive information, outside of my own code, with a full understanding that I may well be sharing my IP with unwanted processes. I do what I can to mitigate this, but I’m sure it’s not enough.
For performance I think Claude or ChatGpt are the best, but it varies depending on your use case. Best way to find out is to subscribe to a service that offers you the choice to alternate between them, to figure out which works best for you.
The company where I’m currently employed lets us use Github Copilot, even though this is a horrible idea in context of the sensitive information they deal with, and the small edge this gives their developers.
I use AI to search for information a lot of the time, and when I’m satisfied with my results I usually cross reference them against other sources. AI is great at putting you on the right track, but still not good enough to give you exactly what you needed most of the time.
As for workflow. I usually create a mental model, covering all the code I want in the way I want it. Sort of like making a map. Then I ask the coding agent to write it piece by piece, not all at once. If there is too much to do, it gets confused and the code becomes a mess.
After it completes each piece, I review it, fix potential problems and then move on to the next piece.
I'm a senior web developer and help maintain hundreds of PHP repositories for work. I avoid the use of AI as best I can. When I do ask questions to an LLM, I feel that I'm partaking in a dirty habit that I should quit. I don't have anything for that installed on my machine/integrated in my IDE. I feel that using LLMs to understand and solve problems is unreliable, a barrier for my personal development, a threat to the future of the industry I work in, unfair to those who wrote the content it is trained on, and bad for the environment.
I also largely feel the same way about other AI products, such as image generation.
Also use them for completion to spec of specific functions/components, as a kind of pairing. Permits staying at high cognitive engagement again without getting bogged down.
A lot of work is local, qwen and mistral models are go tos.
Recently tackled a huge project solo; something large enough with high enough performance that i wouldnt attempt it without a team to build it.
I have a working prototype in production thanks to ai, it's nearly entirely coded with AI. It's big enough of a project that im watching gemini pro take 5 minutes, and fail at replace strings. Which pisses me off because it's burning my limit on its own fails. It's also pretty vague with limits; im always surprised when i finally hit the limit.
Now, what ive achieved would have been utterly impossible for a non-dev to 'vibe code' no matter the quality of AI model.
>If you do, what workflows have worked best for you? What models do you trust? Do you believe that it may be a problem offloading thought you would do yourself otherwise to a model?
I use many different models. I trust them all frankly. GPT 20B is ultra fast and can be trusted. Different models for different purposes. Obviously big refactors, it's preferred to go with a big cloud pro model.
On the backend, I still mostly code by hand, but I consider LLM tab-completion essential and a basic tool. When I use LLMs to generate code more extensively, I am usually pretty specific about what I ask the LLM, modify the code more, question it, etc.