We have this problem with estimation. I think we all do have this problem now as I have not seen “THE solution” yet in the market. Although to be honest with ourselves, it will probably never be one. There was never one for software engineering estimation. Many different techniques and approaches and they were all insufficient. Estimating is hard. It’s hard because humans are non-deterministic. So are AI models.
Frontier AI models are very smart now. More than ever before. For estimating, we had been using a % of discount to account for AI efficiencies. We had been working with 30% as a general number. That number is so inaccurate today that as a result, even though we are accelerating with AI, the estimates look so inflated like if AI never existed.
I believe estimation had to change. Nowadays, it is not so much about estimating what needs to be done (from scratch), instead, it’s about estimating what else is missing, what are the gaps the AI miss or failed to implement. Frontier AI models are so good now that they will get a lot of the work done, with good quality overall. Not perfect but more than acceptable.
I ran a test passing an RFP doc 50 pages long to cursor. I used multitask mode to extract requirements, screenshots and put together an implementation plan. Fuck, I was impressed with how just a one-shot prompt did so much for me. It did it almost like I would have done it manually, very clean and organized. It extracted every single screenshot from the PDF, link it to the requirements. I asked it that EVERYTHING needed to be traced back to the original PDf, and so it did.
Then the implementation was so fast. Composer 2 is impressive. Really impressive. It makes me not wanting to use Opus. Opus is a beast but it is expensive as hell, slow too. With Composer 2 is like magic. It wrote all the code almost instantly.
I love Claude Code and its harness. I love their models. I use them for everything, but damn how good is Cursor in general. I kinda hate their now dual UI with the Editor view and Agents view, it drives me nuts, but hell how easy are they making it to developers to be productive. It’s crazy.