The alleged illusion of AI productivity

This illustration by Forrest Brazeal captures the current state of AI-assisted software engineering, also sometimes referred to as vibe coding, rather well.

I would add though, that there is a middle ground here. Repetitive tasks, front-end code, generally all “linear” tasks (where extrapolation is harmless), can be achieved a lot faster with AI. It’s when you throw the whole repo at an LLM and expect it to do magic that you get those huge spikes and drops. A senior developer, using a team of “junior LLMs” can still get a lot more done a lot faster than before. As the LLMs and the systems connecting them get better, this will improve.

Even though the experience of using LLMs for coding might be somewhat taxing at the moment, it is probably worth the initial cost in exchange for getting to grips with how to work with AI in the loop.