These hype stories all seem to make the same error: confusing writing code with developing useful software.<p>Did spreadsheets on every PC put accountants and CPAs out of business? Did home thermometers and blood pressure monitors put doctors out of business? Of course not. Learning how to use a tool (programming language) represents a necessary but not sufficient part of developing software.<p>ChatGPT <i>et al</i> can generate some good code, often subtly wrong or failing to meet requirements, but impressive nonetheless. That probably shows that most code we ask it to write already exists many times over — a fact of programming that matches my own experience.<p>I asked ChatGPT (free version) to write FizzBuzz. It gave a correct solution. Then I asked it to change the code to count only even numbers from 2 to 100. It got that right too. Then I asked it if that version would ever print ”fizz” or “buzz” and it confidently told me no, because no even numbers are evenly divisible by three. This kind of lapse in reasoning ability is going to cause more havoc than errors in spreadsheets already do, when people who can’t <i>read</i> code, or describe a business requirement unambiguously start deploying code from LLMs.
I think ML is basically like a new programming paradigm and for what it works for it will be great. Not sure that we’re really that close to it being the only paradigm though. As far as the other end, its output still requires editing to put into production. So, while Manjoo is wayyy less fabulist than most mainstream takes, I still think he’s overselling.