I tend to think about it the same way as you. Having an algorithm for multiplying two numbers is qualitatively different than having a lossy compression of a hu

I tend to think about it the same way as you. Having an algorithm for multiplying two numbers is qualitatively different than having a lossy compression of a huge dataset of multiplication tables. The latter is what GPT3 has and it just doesn't scale.