Software was once my line of work. I've lived through too many software revolutions to expect all that much from AI. There were high level languages, structured programming, object oriented programming, automatic memory management, relational databases, sand boxing, code validation, algorithmic verification and at least a dozen others. And, here we are.

(1) AI seems to be focused on the easiest part of coding, cranking out lines of code. The hard parts are figuring out what the software needs to do, dealing with the APIs and their discontents, dealing with obscure and unreported runtime errors, improving performance and eliminating all sorts of little glitches. Writing the code is usually about 20% of the work.

(2) We have more programmers than ever, especially if you count declarative languages like screen scrapers, report generators, HTML and CSS. If AI makes some class of programming easier, odds are that standards will change to demand even more programming. Call it induced demand.

(3) People will be surprised at how much labor it will take to train and validate AI systems. Someone has to gather the corpus, validate it, train the system, verify and tweak it, apply it and update it. Add dealing with regulation some time in the next decade. Since AI algorithms are not transparent, fixing problems will require a lot of skills that barely exist today.

Expand full comment

"Industrial policy? No, I don’t think its first-order effects must be to boost capital and avoid boosting labor."

Certainly not "avoid." But I don't see why each and every policy/policy area has to try to achieve multiple objectives. [Tinbergen]

Expand full comment

The first Ken White link takes me to Gmail

Expand full comment