I agree with the sentiment, but I don't think the example given (creating SQL queries) is a good representation of this problem.
That's because if you know a little bit of SQL and know how to validate the answer LLMs give you this becomes a non-issue.
A better example would be an ambiguous prompt where the LLM can either use an array or a map to solve your problem so it chooses an array. But down the road you need a feature where direct access is what you need, but your code is already using arrays. In this situation what tends to happen is the LLMs ends up making some hack on top of your arrays to create this new feature and the code gets real bad.
By not understanding the difference between these 2 data structures you aren't able to specify what needs to be done and the LLM ends up implementing your feature in an additive way. And when you add enough features in this way things get messy.
What is still not clear to me is what is the proper "abstraction layer" we need to use to learn things in this new world where LLMs are available.
That's because if you know a little bit of SQL and know how to validate the answer LLMs give you this becomes a non-issue.
A better example would be an ambiguous prompt where the LLM can either use an array or a map to solve your problem so it chooses an array. But down the road you need a feature where direct access is what you need, but your code is already using arrays. In this situation what tends to happen is the LLMs ends up making some hack on top of your arrays to create this new feature and the code gets real bad.
By not understanding the difference between these 2 data structures you aren't able to specify what needs to be done and the LLM ends up implementing your feature in an additive way. And when you add enough features in this way things get messy.
What is still not clear to me is what is the proper "abstraction layer" we need to use to learn things in this new world where LLMs are available.