AI usage has settled back into the "99% easy, 1% hard" rhythm.
When you give AI full permissions, working in a domain you know well, you can let it run for three or four hours and build the thing out. But the moment you enter unfamiliar territory, you need to keep prompting, and iteration slows down. And when running on a personal machine, it's hard to truly delegate — a lot of time goes to negotiating instructions with the AI.
The productivity boost is real, but it varies enormously across scenarios:
- Familiar domains: Roughly 10x efficiency gain.
- Completely unfamiliar domains (0 to 1): Paradoxically, maybe 10–30x, because AI lets you skip the massive onboarding cost.
- The middle ground: Some knowledge but not much, domains that require real domain expertise. Roughly 2–3x.
The third category is the hardest. Hard because you don't know how to define the end result, don't know what to expect, and don't know how to set up a framework where AI can self-run and self-evaluate. The result: tons of time spent going back and forth, letting AI educate you while you rapidly build your own evaluation framework during that education, then turning around and describing it back to the AI.
About two-thirds of the time goes into this educate-and-align process. Hopefully it's a one-time investment, but I'm not sure.
And in an era where AI fully empowers you, it's easy to lose focus. You keep entering new industries, hitting the same bottleneck every time: you need to be educated before you can effectively give direction.
This raises a question: should there be a barrier? Even with AI assistance, maybe some domains still shouldn't be handled by me — better to find someone with deeper expertise in that area.
How to measure that ROI — still figuring it out.