Very much doubt #1, the whole reason this works is model size. Smaller models and some intent model on top of it are very unlikely to produce the results they'v

Very much doubt #1, the whole reason this works is model size. Smaller models and some intent model on top of it are very unlikely to produce the results they've been able to (and goes against everything they've shared on how this works). For #2, they've talked about this openly (and if you use any of the underlying models directly you can see it), there is a limited context window (~4k tokens for davinci-003), so time aside, they have to compress to fit that context window.