Automating ML Research to Improve Sample Efficiency

The labeling side of that is not so new, but you're thinking we could combine that with some ability to generate or make better use of that data. And as it improves and gets smaller, human has to do fewer bad examples for every good one. You expect this to happen around the same time as sort of early week AGI. Yeah. There's a lot of low hanging fruit here with lots of different architectures that haven't been tried yet.