80% Pleasure and 20% Suffering

Today I was talking with my colleagues about using AI in new projects, and I told them that every project is 80% pleasure and 20% pure suffering.

It is something we go through constantly whenever we build a new tool. We usually use SDD for complex tools and vibe coding for simple ones, and either way the result is the same: we start out thrilled by how fast the project is moving, how well it captures our ideas, how it works and the amazing things we get done in so little time. And so it goes until we reach the 20%.

That 20% is when the product is practically finished, but you start fine-tuning, going over bugs, implementing the new ideas… and that is when, holy hell, everything starts to get complicated. The AI doesn't understand you, you get frustrated, it fixes one thing and breaks another… and you say: “WTF!!!

If the tests pass (the unit tests, the integration tests, the e2e tests), why are you breaking it?”. You watch what looked like a matter of 6 or 8 days drag on (“damn, I've been at this for 2 weeks, it should be finished by now, and it still isn't running as smoothly as I want”). Then you start questioning everything: “I'm useless”, “this AI is crap”, “maybe I'm explaining it badly?”.

That is when you turn to “adversarial networks”: basically, using another AI to write the tests (works wonders, by the way), and even better if you take away its knowledge of the code. Little by little you fine-tune it and deliver the product, but you look terrible in front of the stakeholders: you are 2 weeks late.

Over time we are learning to shrink that 20%, but since the models change every five minutes, we have to keep redefining how we work and adapting to the new capabilities AI offers.

Off-topic: we still don't know why some colleagues burn through their tokens in half a week while others reach the end of the week with quota to spare… That is worth a study of its own.