It became too easy to delegate all my software work to the LLM. The work feels faster and easier but makes me less aware of the code I'm responsible for.
The LLM is fine until it eventually makes the mistakes. Stepping in to such generated code will cause me problems understanding it. The LLM doesn't reduce the entropy, it increases it. It makes spaghetti. Trying to untangle it is a surefire way to exceed the project's budget.
It's worse for my focus, too. When working on problem myself I will work slowly, but smoothly. When using an LLM to think for me I get frustrated more easily in the constant prompting loop, especially when the model gets it wrong again.
Using LLMs is not inherently wrong. It gives an advantage even if I don't ask it to write whole features for me. However it works better when I let it criticize or follow my own thoughts rather than completely lean on it. Asking it to write long parts of code is cool, but I need to control how it does it.
As specialists we need to keep operating on the top level of thinking. This is how we don't get replaced by the LLMs. The LLM should be an assistant, not the driver. It should criticize and obey, not create. Let's not give in to the pressure for working rapidly and mindlessly.