[sublinear]

I think use cases like this are the right ones for generated code from language models.

I'm pretty sure the vast majority of training data is more buggy than not and literally writing the code is only half of what it takes. The other half is comprehension of the problem it solves which is usually missing from documentation at the level of mere snippets.

[908B64B197]

Something I wonder is what the computing landscape (and tech recruiting) will looks like once we reach a critical mass of "GitHub Co-Pilot native" devs. Especially with the bootcamp craze.

Are we going to see hires basically trying code generation until it builds?

[groestl]

Take something perfect and internetlify it. It was basically trained on exactly that data.