Hacker News Reader

What kinds of problems are you trying to solve that can be solved by short program tapes but aren't obvious to a human?

I have a hypothesis that the smallest possible program that can handle things like natural language translation will be much smaller than we are expecting it to be.

Anything where we are transforming information from one format to another on a ~1:1 basis without injecting additional external information (i.e., what foundation models do with their broad world knowledge).

I don't know what your expectation on the program's size is, but even for a constrained translation case (i.e. legal transcription between two Romance languages), I would expect you are embedding at least the size of WordNet in both languages.

Data modelling, especially in high dimensions can be non-obvious to a human, yet a tractably small program can do well.

Reply on news.ycombinator.com