I think LLM is fine for shorter scripts. As a professional programmer, it has helped me with writing simple throwaway scripts. Those circumstances are rare.
My stance is that if you think LLM help you get your job done, then use LLM. It’s just another tool to your arsenal.
I don’t trust using LLM for large long running software projects though.
I have been building various things with AI coding tools for a month or so now. I rate the various engines on how far I can take them before they get hopelessly lost, unable to correct their own errors. For the best tools this seems to come after about 50 to 70 iterations of asking for small feature additions or error corrections, weaker tools (like Copilot) hit these infinite loops of fixing their errors with other errors much faster.
It’s a good limit, because after 2-3 hours of AI interactive development, I can then spend 4-6 hours going through the resulting code - cleaning it up and understanding how it works. I suspect if AI were taking me farther, like 100-150 iterations, it would probably take me more like 15-20 hours to unravel the various things it comes up with - kind of a point of diminishing returns.
Bottom line: think of your project in terms of microservices. AI is pretty good at microservices. As long as the individual services are each robust in their delivery of the required functions, you’re in good shape.
If it ever becomes “mystery meat,” it’s time to recode by hand.
I think LLM is fine for shorter scripts. As a professional programmer, it has helped me with writing simple throwaway scripts. Those circumstances are rare.
My stance is that if you think LLM help you get your job done, then use LLM. It’s just another tool to your arsenal.
I don’t trust using LLM for large long running software projects though.
I have been building various things with AI coding tools for a month or so now. I rate the various engines on how far I can take them before they get hopelessly lost, unable to correct their own errors. For the best tools this seems to come after about 50 to 70 iterations of asking for small feature additions or error corrections, weaker tools (like Copilot) hit these infinite loops of fixing their errors with other errors much faster.
It’s a good limit, because after 2-3 hours of AI interactive development, I can then spend 4-6 hours going through the resulting code - cleaning it up and understanding how it works. I suspect if AI were taking me farther, like 100-150 iterations, it would probably take me more like 15-20 hours to unravel the various things it comes up with - kind of a point of diminishing returns.
Bottom line: think of your project in terms of microservices. AI is pretty good at microservices. As long as the individual services are each robust in their delivery of the required functions, you’re in good shape.
If it ever becomes “mystery meat,” it’s time to recode by hand.