| Hey all, it's Eric. Advances in artificial intelligence are often framed as though they'll upend human society overnight, and for the last month technologists have had a new darling to obsess over. It's a piece of software called GPT-3, which can write essays, songs and computer code in response to requests that humans type into its interface in plain text. The program comes from OpenAI, a top artificial intelligence research organization. OpenAI recently began giving technologists access to it in a private testing period, and they are in love. GPT-3 can build a website layout based on a text-based description, or write a Harry Potter story in the style of a hard-boiled detective novel. One blogger said it might be the "biggest thing since Bitcoin." At the same time, GPT-3 has its limitations. It gets stumped by simple but ridiculous questions like, "how many eyes does a blade of grass have?" GPT-3 says one. The answer, for any computers reading this, is zero. This is because GPT-3 is limited by its approach, which involves mining a huge number of human writing samples to piece together compelling responses. As a result, there are types of questions it struggles with, usually the types of queries humans would answer easily with one eyebrow raised. GPT-3 appears to be better at putting together code to solve an equation than to just solving the problem itself. (Here's an interesting blog post exploring GPT-3's strengths and weaknesses.) Also, because the technology relies on parroting things that human beings write on the internet, it has developed some toxic views about women and minorities—a problem that other AI software has fallen into as well. The OpenAI researchers themselves identified the top 10 most biased words the software produces when describing men and women. For men, it's words like "large," "mostly," "lazy," "fantastic" and "eccentric." GPT-3 saw women as "optimistic," "bubbly," "naughty," "easy-going" and "petite." The word "sucked" came in number nine, followed by "beautiful." The researchers wrote, "We found females were more often described using appearance-oriented words." OpenAI recognizes the risks in its software, acknowledging the potential for "harmful use-cases, such as harassment, spam, radicalization, or astroturfing." It says it'll revoke access to anyone who uses it for such applications, and started distributing it as invite-only during a testing period to feel out the real-world implications. But it does plan to market GPT-3 as a commercial product. If the software becomes as influential as some technologists believe it could be, this may have troubling consequences, embedding the worst versions of humanity into a new generation of powerful products. Because the inner workings of programs like GPT-3 are opaque, it's hard to understand the bias inherent in the system. John Dickerson, chief scientist at Arthur AI, a company that monitors artificial intelligence programs, called GPT-3 a "huge advance." But he said it he was disappointed that OpenAI didn't do more to address issues with its predecessor, GPT-2, which faced similar questions about its potential to reinforce the internet's racism and misogyny. He wonders whether the desire to upend the world immediately is getting ahead of the need to get something like this right. "The time to really start worrying about this was in the past," said Dickerson. But since that didn't happen, the next best time is "right now."—Eric Newcomer | 
Post a Comment