I agree largely, but hold up a bit on the statement that getting an LLM to write the code means I never solved the problem. That may be true for a lot of vibe-coders - loops of "write this code" "no, not that", "fix plz", etc. But what I've found is that Claude (at least) does best on small problems that are very well defined - in other words, I have to solve the problem _before_ I can get Claude to produce the code for me, if for no other reason than that the model will model along, and needs an expert on the problem to be able to push back when it's wrong.
So I agree that leaving an LLM churning for a week or two and then claiming that you have a product to sell is tenuous, but I disagree that one can't both use an LLM _and_ understand the solution - it just takes active participation towards that result.
Whilst there's no remuneration involved, I view projects I put up on github (that I also tell other people about) in the same light - so, for example, the largely-ai-written https://github.com/fluffynuts/nestray, whilst designed to scratch an itch of mine, may be useful for others because I've aimed at some level of genericism. Is it a perfect product? NO! But it does what I want, and probably what someone else might want with respect to showing email notifications without requiring any account access (simply reads from local Thunderbird files)
Recognising this has also made me recognise the real problem of using an LLM without expert knowledge in the domain - if you don't know when it's wrong, you won't know when to push back. So whilst I can be very critical of code outputs, for example, my knowledge of quantum mechanics is so rudimentary that it's practically impossible to push back on conversations about QM. So who knows if my hobby knowledge actually holds any value? The best I can do here, when repeating the information, is to say "at least, that's what the machine says".
I agree largely, but hold up a bit on the statement that getting an LLM to write the code means I never solved the problem. That may be true for a lot of vibe-coders - loops of "write this code" "no, not that", "fix plz", etc. But what I've found is that Claude (at least) does best on small problems that are very well defined - in other words, I have to solve the problem _before_ I can get Claude to produce the code for me, if for no other reason than that the model will model along, and needs an expert on the problem to be able to push back when it's wrong.
So I agree that leaving an LLM churning for a week or two and then claiming that you have a product to sell is tenuous, but I disagree that one can't both use an LLM _and_ understand the solution - it just takes active participation towards that result.
Whilst there's no remuneration involved, I view projects I put up on github (that I also tell other people about) in the same light - so, for example, the largely-ai-written https://github.com/fluffynuts/nestray, whilst designed to scratch an itch of mine, may be useful for others because I've aimed at some level of genericism. Is it a perfect product? NO! But it does what I want, and probably what someone else might want with respect to showing email notifications without requiring any account access (simply reads from local Thunderbird files)
Recognising this has also made me recognise the real problem of using an LLM without expert knowledge in the domain - if you don't know when it's wrong, you won't know when to push back. So whilst I can be very critical of code outputs, for example, my knowledge of quantum mechanics is so rudimentary that it's practically impossible to push back on conversations about QM. So who knows if my hobby knowledge actually holds any value? The best I can do here, when repeating the information, is to say "at least, that's what the machine says".