Read on Twitter
Thanks, @ChrisGPotts for this thoughtful essay! As someone who has published what I believe *is* "compelling reasons to think they can't", and in the spirit of continuing the discussion, I’d like to respond here. https://medium.com/@ChrisGPotts/is-it-possible-for-language-models-to-achieve-language-understanding-81df45082ee2
My first question to you is about your definition/operationalization of "understanding": "Is it possible for language models to achieve truly robust and general capabilities to answer questions, reason with language, and translate between languages?" >>
Why do you find this a satisfying definition of "understanding"? Also, how do you detect "reason[ing] with language" as distinct from manipulating form, but without doing any reasoning?
My second question is: Do you think that calling the above "understanding" is perhaps an overclaim? How would you reply to someone who thinks it feeds into AI hype?
Finally, regarding "we would be hard-pressed to define necessary and sufficient inputs for successful human language learning", I think there is a clear necessary and sufficient input: Joint attention. (Which can be achieved with different modalities.)
