First, there's the uselessness of entry-level engineers. They are hired for their growth potential (that every human has, but language models don't) and are expected to grow into mid-levels and seniors. In most of the sw industry, entry-level roles are also not terminal, which means that such an engineer must get promoted at least +1L to keep working in the company. Someone (or something) perpetually stuck at the entry level is a bad value proposition for a sw company.

Secondly, the code that language models produce is buggy. They can, on occasion, produce amazing code and even entire codebases. But this is an exception, not the rule. You can generally prototype something or get an idea for something from language models, but you can only push very little of that into production. What good is code you can't use? You still need an engineer to oversee the language model's outputs.

Overall, if some company replaced their juniors with AI, that would be incompetent management.