This is an introduction to a paper, the full text of which is found here. Michael Feldstein has alo provided some AI structures that both help explain what the paper says and test the predictions offered in the paper. I didn't use the AI components, but I did read the intro post and the paper as a whole, which was well worth the effort. It defies summarization in a short post such as this, but here goes: transformer-based AI (such as ChatGPT) learn complex and apparently rule-based systems (such as language or chess) by preserving distinctions that have predictive import in a given context, and discarding the rest. Feldstein calls this the conservation of predictive meaning (CPM) theory. My assessment is that he is not wrong. I say it that way because I would word things a bit differently and draw slightly different conclusions. What he calls 'distinctions' I would call 'patterns'. What he calls "a general mechanism to reduce predictive surprise" I would call 'salience'. I would not say language learners are "like effective cryptographers", nor would I say they "decode what has been communicated." Overall, though, I think he is on the right track.
Today: Total: [Direct link] [Share]
