by Utsa Chaturvedi
In February 2016, 29 paintings created by Google’s artificial intelligence were sold at an auction called “DeepDream: The art of neural networks” in San Francisco, the most expensive artwork sold for $8000. Yes, AI can create art, but that’s not all. Writing songs, stories, and, more recently, software, are just some of the many skills that AI has acquired.
AI essentially leverages computers to mimic the problem-solving capabilities of humans. Through the rise of systems called neural networks, machines can learn skills by analyzing vast data sets. From Netflix recommendations to email spam filters, from training self-driving cars to medical imaging that identifies lymph nodes, AI is an expansive field of research, and its impact on our lives is significant. But what are the implications of software writing its own software, and more importantly, for the programmers who taught the AI to write its own code?
Recently, researchers started designing neural networks that analyzed extensive amounts of text on the internet. The aim was to train networks to predict words in a specific order. GPT-3, or the Generative Pre-Trained Transformer 3, is one example of a language model, developed by OpenAI, an AI laboratory based in San Francisco. When prompted by small quantities of text, it can produce prose written in a style similar to the prompt, and write its own Twitter posts, speeches, poetry, and news articles. Moreover, it can be used to write code and is already being used by firms like Microsoft, who invested $1 billion in OpenAI last year and gained exclusive licensing rights to GPT-3. “If you can describe what you want to do in natural language, GPT-3 will generate a list of the most relevant formulas for you to choose from,” said Microsoft CEO Satya Nadella. “The code writes itself.”
Using the same methodology as it does when producing text, GPT-3, when fed vast amounts of pre-existing code, can be trained to predict the code a programmer needs next. As a programmer types, potential “code completions” might pop up. Seeing the value of such developments, OpenAI created Codex, a new AI technology that writes computer programs. They trained it on a vast array of code, and what Tom Smith, who runs an artificial intelligence start-up, found when he tested it was no small feat to accomplish. Codex was given a job interview and asked to tackle “coding challenges” that programmers interviewing at firms like Google face, like writing a program that replaced all spaces in a sentence with dashes or identified invalid ZIP codes. It completed both successfully, in addition to other tasks that would be daunting for human programmers, and being able to translate from one programming language to another in 12 languages.
“These are problems that would be tough for a lot of humans to solve, myself included, and it would type out the response in two seconds. It was spooky to watch,” he said.
While self-coding AI does have positive effects, it is important to consider its impact on humans. However, research shows that instead of replacing humans, this technology will only help make their work more efficient. Models like Codex are not fully accurate, based on code written by humans, which itself is not completely reliable. While its skills are impressive, Codex can only mimic what it has seen before, not think on its own. Often, the programs generated by OpenAI don’t run, contain security flaws, or produce a different result from the desired one - OpenAI estimates that Codex produces the right code 37 percent of the time. Founder of AI lab Fast.ai Jeremy Howard says, “It is a way of getting code written without having to write as much code. It is not always correct, but it is just close enough.” In addition, GPT-3 has generated offensive text about Black people, women, and Muslims - it shares many human prejudices, and until OpenAI can find ways to address this, GPT-3’s use could be limited, yet another reason why humans are needed to review content produced by technology.
In essence, while the development of self-programming AI is useful to human programmers, it will be a long while before it can replace humans, if ever. After all, however close AI gets to recreating Van Gogh’s art, it is doubtful that it will ever achieve the level of emotion and proficiency that he did.
Comments