Can Artificial Intelligence augment humans in coding?

Post Reply New Topic

STORY: A.I. Can Now Write Its Own Computer Code. That’s Good News for Humans.- A new technology called Codex generates programs in 12 coding languages and even translates between them. But it is not a threat to professional programmers.

https://www.nytimes.com/2021/09/09/tech ... oding.html

I don't think this poses any threat to coding. Sure, basic algorithms could be auto-generated and a lot of the boilerplate code can be automated, which would only give coders a lot more capability to do more advanced coding on top. If anything it would increase demand for software engineers, not decrease it. One of the goals of software development for decades has been to make coding more modular. If we had ways to connect existing systems and build on top of them easily, then we wouldn't need to write software from the ground up any more. Today we have libraries and frameworks to fill this role, but there's always room for improvement.

But here's why it won't replace human coders. Stuff like this AI-generated code can take care of a lot of the mundane aspects of coding that have been done thousands of times before. But it still can't generate novel code. It requires a lot of existing human-written code in order to train the models. And it can't create anything it hasn't seen before, because it doesn't actually understand the code. It's just applying patterns and matching existing code to new scenarios. Maybe it can infer programming rules etc and construct new code to match a new situation. But it can't solve problems it doesn't know about, and you can't just tell it to build you a backend + frontend that can run on AWS and scale to millions of users without being too expensive, or refactor an existing piece of code to add a new feature. It doesn't even know whether the code it suggests will do the thing you asked for. Only a human (currently) can verify that a piece of code does what you want. I don't think that will necessarily always be the case, but we're currently a long way from AI understanding code.

In my opinion coding will be one of the last things to be taken over by AI. Once a computer can write its own code, it's game over for humanity. If an AI can code at human-level or better, then it can code a better AI. Rinse, repeat. Now you have an AI that can outsmart humans at everything and if it wants to control the world it can manipulate humans into doing anything it needs. Such an AI would also be smart enough not to let humans know it had this capability until it was far too late. Of course there are many unknowns so it's virtually impossible to predict what would happen in such scenarios.

What we'll see in the short term is simply automation of repetitive coding tasks, and that's exactly what Codex does. And that's a good thing for programmers, because we can focus on more interesting things and build software faster.

Another thing, if AI can code, and by extension can code a better AI, then it could program / build robots to perform literally any other job. So this is why I think coding will be one of the last things AI takes over from humans.

Perhaps - but probably not now. When I was young a computer could beat a average chess player at chess, but could never beat a grand master. There was a theory that there was something special about a grand master and a computer would never be able to beat a grand master. Now there are computers that can beat grand masters at chess. Computers can also bead human experts at Poker and Jeopardy ( a TV quiz or trivia question game ). Some day a computer may be able to write code better than a human - or write some code while a computer guides it. Someday a self-driving car may be a much better driver than the average human driver. But not today. Part of writing good code is understanding what is wanted. Understanding what is wanted - in general - is probably AI Hard - that is just as complicated at solving the "thinks just like a human" problem. I would think there are some types of things that require coding now that will only need a computer in the future.
the LISP programming language is particularly good for AI programming because writing code that writes code is easy and normal. In Java it is strange and abnormal to write code that writes code. But you can always write a LISP interpreter in Java and then you are off to the races. There are also special modules that let you, for example, implement an abstract class at run time or generate JVM compatible byte-code class files which can then be loaded at run-time. Most programming languages can create binary files and a binary file with just the right bytes can be an executable file - and most programming languages can start an executable file. It is just a matter of time, energy and money.

Great things have been happening in natural language lately. It is foolish to think that more advances will never come.

Codex's output is often just terrible... and uncreative.