Hypothesis: the actual long-term impact of LLM's on coding will be that more different languages get used, because it reduces the "ugh I don't want to have to learn all that" psychological obstacle that prevents programmers experienced in one language from learning another. We should be using different languages for different tasks, but the push has always been to try to get One Language To Rule Them All. Ten years ago people tried to make Javascript their db query language and their backed language. More recently, the push is to use Python for everything. Both languages are fine in certain domains, but a poor fit in others.
Even if the idea that an LLM will help you do it is false, perhaps it is still a good idea if it convinces the experienced programmer to go ahead and use SQL for the query, Go for the async, Javascript for the frontend, etc. Right now, few if any companies would let you use the best tool for the job, if that's not one they already use. Perhaps the best use of LLMs is to convince programmers, and their bosses, to use the best tool for each job.
But, after you've gotten past that part, you will probably (like the author of this article) need to throw away the LLM-generated code and write it yourself.
I have a hypothesis that at some point someone will do a study of "which language do LLMs work best with?" and we will find out some languages and frameworks are better for LLMs than others. If we stick with LLMs for a while we may even develop languages and frameworks specifically to be easy for LLMs to generate code with. And then people will architect their software to use the approach that is easy for LLMs
counter hypothesis - fewer languages will be used because LLMs will be better at languages with lots of training data
I never write python in my day to day but every from-scratch project I've knocked out with Claude code has been in python because that's what it seems to default to if I don't specify anything
Interesting point!
I wonder if this will also mean that new languages (or even algorithms or code patterns) are harder to get adopted, because the mass of existing code (that LLMs learned from) exerts a gravitational force pulling things back down to the status quo.