Stephen Downes

This article discusses abstractions for machine learning specifically, but the argument generalizes. Most of computer science consists of abstractions. At the base level is the machine language that is the direct interface with the hardware. But because this is so complex, more abstract languages are built on top of it (most notably the computer language 'C'). But even this is pretty complex, so above C even more abstractions are built: languages like Python or Perl, which are based on C compilers. But with each abstraction, there is a trade-off: clarity versus efficiency. But if AI, not humans, are writing the code, we don't need the abstractions. "What do we keep? The intent, the invariants, the tests, and the hard-earned domain knowledge that currently lives inside ThunderKittens's abstractions and keeps them correct on hardware."