Yes, our job as programmers includes much more than just writing code, for example mentoring young talent and developing high-level systems. But programming has always been at the heart of our work. Throughout my career, I have been interviewed and selected precisely based on my ability to solve small programming puzzles. Suddenly this ability lost its meaning.
This is what Ben (a friend of the author) told me, who kept telling me about the spectacular successes he had achieved with GPT-4. It turned out that he was not only good at tinkering, but also had the qualities of an experienced engineer: he had extensive knowledge and could suggest solutions to a problem. For one project, Ben wired a small speaker and a red LED light bulb to create a portrait of King Charles, with the light bulb representing the jewel in his crown. The idea was that when you typed a message on a companion website, the speaker would play a tune and the light bulb would flash the message in Morse code. Ben couldn’t program the device to retrieve new messages. It seemed to require expertise not only in the microcontroller he was using, but also in Firebase, the back-end server technology that stored the messages. Ben asked me for advice and I mumbled a few options. In fact, I wasn’t sure if what he wanted was possible. Then he asked about GPT-4. He told her that Firebase had a feature that would make the project much easier. Here it is, and here is some code to use that would be compatible with the microcontroller.
Afraid of the idea of using GPT-4 myself—and feeling a little uncomfortable about having to pay OpenAI twenty dollars a month to do it—I nevertheless started trying out its capabilities through Well. We sat down to work on our crossword puzzle project and I said to him, “Why don’t you try arranging it like this?” He offers me the keyboard. “No, you’re in charge,” I told him. Together we got an idea of what AI could do. Ben, who had more experience than me, seemed to be able to make the best of it in the event of a stroke. As he later said, his own neural network had begun to adapt to GPT-4’s. I would have said he had achieved mechanical compassion. Once, which particularly amazed me, he asked the AI to build him a Snake game like the ones found on old Nokia phones. Then, after a brief exchange with GPT-4, he got the AI to modify the game so that when you lose, it tells you how far you’ve deviated from the most efficient route. The robot needed about ten seconds to do this. It’s a task that I honestly wasn’t sure I could handle on my own.
In chess, which has been dominated by AI for decades, a player’s only hope is to cooperate with a robot. These half-human, half-AI teams, called centaurs, could still beat the best humans and the best AI machines working alone. Programming has not yet taken the path to failure. But the centaurs have arrived. GPT-4 alone is a worse programmer than me at the moment. Ben is much worse. But Ben plus GPT-4 is a dangerous thing.