Okay, I really didn’t want to add yet another post about LLMs and programming to an already tiresome sea of existing discourse, but wrote this as a comment on a LinkedIn post and thought it might be worth sharing here. The topic was about people posting hyped-up claims of 10x (or greater) productivity boosts by having LLMs write their code.

I find the 10x (or even 100x!) claims ridiculous. Even if LLMs enabled the code to be written at infinite speed and perfectly correct (which of course, it cannot), there would still be a massive problem:

If we exclude 1-person startups, then it probably ranges from 80% at most, down to 30% or below for process/meeting-heavy teams. Much of the time is spent on activities that can’t be sped up with LLMs, like planning, alignment meetings, discussing requirements with the customer etc.

Since less than 100% of the time is spent programming, this means Amdahl’s Law applies. That is, if 50% of time is spent on activities that can’t be sped up (meetings, thinking, reading requirements/design docs), then LLMs can provide AT BEST a 2x speedup, assuming they produce correct code instantly. Even if you would otherwise spend 80% of your time coding, LLMs can deliver at best 5x.

I would suggest that anyone claiming 10x (or even, honestly, 2x) benefits is lying, unless they’ve eliminated all meetings and someone hands them perfectly clear requirements which they feed into a magical always-correct LLM, which would suggest that human serves no function anyway.