Vibe coding is the new product management. Training and tuning models is the new coding.

You might demur, it’s one of those pithy Naval sayings, but it is food for thought nonetheless — if the LLMs are failing in your domain, you wait for the model to get better, OR you pull up your sleeves and see if you can improve the model by tuning it.

Can you just not fix the broken code generated by LLM? you ask. The keyword here is “vibecoding”, which implies you are generating most of your code without going through the LLM’s output line by line. So, in order to figure out what’s wrong, you have to read it all - which is putatively against the vibecoding ethos.

If you consider coding agents to be your assistant, you want the assistant to get better over time without you having to step in and fix the messes all the time.

Right now, tuning is quite the esoteric art. Perhaps, the next jump in building trust with coding agents will come from making the tuning process tractable for working programmers.

While there are, and will be, many programmers who will never trust LLM to do what they do, developing confidence in guiding the LLM to make improvements will take vibecoding from “good for throwaway code / good for demo code”, to something broader in scope for developers that have an open mind about use of LLMs in coding.