Natural language programming AIs are helping professional software developers code faster

"Learn how to code." Every time media layoffs are reported, that three-word derogatory phrase is constantly in the mouths and at the fingertips of internet trolls and tech bros. It's a pointless sentiment in and of itself, but with the current development of code-generating AIs, knowing Python inside and out may soon be about as valuable as speaking a dead language with ease. In fact, by taking care of a lot of the tedious programming work, these genAIs are already assisting professional software engineers in coding more quickly and efficiently.

How is Coding Done?

Java and Python are the most frequently used and written coding languages today. When it was first published in the middle of the 1990s, the former nearly single-handedly revolutionized cross-platform operation. It currently powers "everything from smartcards to space vehicles," as Java Magazine put it in 2020, in addition to Minecraft's whole world and Wikipedia's search engine. The latter is the underlying code for many contemporary programs, like Dropbox, Spotify, and Instagram, predating Java by a few years.

Because Java must first be compiled (having its human-readable code converted into machine code that a computer can execute), they operate very differently. On the other hand, Python runs without needing to be first compiled because it is an interpreted language, meaning that its human code is transformed into machine code line-by-line as the program is executed. In contrast to compiled code, which is frequently targeted to a particular processor type, interpretation-based code can be created more efficiently for various systems. The process of writing the code is essentially the same for the two, regardless of how they function: Someone has to sit down, open a text editor or Integrated Development Environment (IDE), and create all those lines of directives. Furthermore, that someone was often a human until recently.

A software engineer will take a problem, break it into several more minor issues, write code to solve each more minor problem, and then repeatedly debug and recompile the code until it runs. This is how "classical programming" is written today. Conversely, "automatic programming" places the programmer at a greater distance. The person produces a high-level abstraction of the task for the computer to generate low-level code instead of writing each line by hand. This is different from "interactive" programming, which lets you create programs currently in operation.

Today's conversational AI coding systems eliminate the programmer by disguising the coding procedure behind a guise of natural language, as shown in Github's Copilot or OpenAI's ChatGPT. The AI can automatically generate the necessary code if the programmer provides instructions on what and how to be programmed.

Codex, an early example of this new generation of conversational coding AIs, was created by OpenAI and launched in the latter half of 2021. By this time, OpenAI had already implemented GPT-3, a big language model that is extremely good at mimicking human speech and writing after being trained on billions of words from the public web. GPT-3 is the predecessor to GPT-3.5, which powers BingChat public. The business then created Codex by fine-tuning that model using over 100 gigabytes of GitHub data. It can translate between existing programs and generate code in 12 different languages.

AlphaCode was created expressly to address these issues by Google's DeepMind. Like Codex, AlphaCode was trained on numerous terabytes of already-existing GitHub code archives before being fed thousands of coding problems from online programming contests, such as counting the number of binary strings of a specific length that don't include consecutive zeroes.