Precision

Precision matters. Or why English is NOT the new programming language.

I really like Erik Brynjolfsson's work and thinking. This figure from his 2022 article in the American Academy of Arts & Sciences Daedalus journal ("The Turing Trap: The Promise & Peril of Human-Like Artificial Intelligence") struck a chord when I first saw it a couple of years ago. I thought "YES!"

With any new technology, especially such a transformative technology as "modern AI", the first instinct of the early adopters is to solve existing problems or improving existing tasks with it. Not to downplay the value of a 10x or even 100x cost or time reduction (which can suddenly unlock vast previously frustrated activities), but the real value usually comes from entirely new "use cases".

But the picture can be misleading in having us believe that the "new tasks that humans can do with the help of machines" fit nicely in the known space of possibilities. I understand the need to simplify the illustration but let's just keep in mind that there are things that come into existence only when they are possible.

However I think something deeper is missing from this picture: the unknown (and perhaps unknowable) dimensions of alien intelligence, the emergence of capabilities that cannot be reduced to human understanding. With some effort, we can assign some "mechanistic interpretability" veneer to that alien intelligence but it remains fundamentally beyond our reach. Which is not to say that it is useless, quite the opposite in fact: it may be how we cure all disease, live to 200 or reach distant galaxies, all things that probably require concepts that we don't even know can be formed. Current models, "next-token-prediction" as they may be, are beginning to show elements of it.