François CholletVerified account

Statistical learning is mimetic, and thus cannot generalize far beyond what it has observed so far. But the human mind can.

A statistical machine translation algo trained on data up to 1800 could not translate a Tom Clancy book. But a smart human from 1800 could.

@fchollet could a current human translate a book written 200 years from now when society has changed so much ? I'm not so sure.

@cbonnett sure. It's not very different from translating old books.

@fchollet maybe,I (maybe naively) think that conveying concepts like cyber attacks would be too different for someone living 200 years ago

@cbonnett given enough text information we can understand any alien concept (e.g. sci-fi). But MT does not rely on understanding concepts

@fchollet @cbonnett Human translated writings at least as old as from 3rd millenium BC. W/o doubt we'd translate text from 200y from now

@glouppe @fchollet could someone from 200 years ago translate nisp proceeding as an extreme example ? Ide argue they couldn't.