Let's predict the future: Within 3 months you will be able to build much more, much faster, with AI. Beyond 3 months I suspect it will get even better, but 3 months is far enough out for our purposes. I recently used a side project to assess the state of the art. I built a native macOS transcription app. It listens to my voice, transcribes it, and pastes the text into whatever app i'm using. There are other apps that do this, but I thought it would be a good candidate for an app built entirely with AI. - it's local. no remote servers involved - it requires a stack i'm not familiar with (Swift, macOS) - it's well-scoped. listen to audio, transcribe it, paste it And of course it's also useful. I have been using Superwhisper for some time, and this app (which I named Heyo) does much the same thing. Beyond doing the transcription, the app also stores the recordings to disk and saves metadata in a database. It makes all this information available via a UI. I knew nothing about Swift UI going into this project. However, by giving clear requirements to o1 pro I was able to build a full, functional app. This is a pretty amazing result. I had tried building this same app many months ago and the project completely stalled out using Claude Sonnet[^1]. However, having recently signed up for OpenAI's $200/mo plan I wanted to put it through its paces. Initially it was a disappointment (see [[A review of o1 Pro so far]]) but after putting significantly more effort into the prompting process I was able to get impressive results. [^1]: It may not have been Sonnet's fault per se. I put much more effort into the prompting this time around and didn't rely on Cursor's built-in context. All context, tens of thousands of tokens, was built up manually. It's possible Sonnet would have done better if I had given it a similar amount of attention. So, if AI is so good now why procrastinate? In short, because its going to get better, and soon. There are two concrete events on the near horizon to look forward to: - OpenAI's o3-mini relesae - Google's stable release of Gemini 2.0 o3-mini is a complete unknown, but considering it's OpenAI's next generation model it may be superior to o1-pro, or at the very least faster which would also be a significant improvement. Google's Gemini 2.0 Advanced model is also quite good in my testing, however, it suffers from a limit on token output. My hope is that this is only due to it being in beta, and that once it stabalizes it will be able to output longer responses. I tried using Gemini 2.0 alongside o1-pro for comparison but the responses kept getting truncated. This meant the entire prompt was unusable, for now. So here's a bet I would take: Whatever you're building, it will take you less time to build in 3 months than it does today. If you can save time by waiting, why not wait?