GPT-3 API is still the hardest single project I've worked on. Most common startup mistake is to start with a technology, rather than a user problem to solve — but working on one problem would sacrifice AGI focus. So felt like it probably wouldn't work, but had to push through.