Write a symphony that evokes an astronaut cat purring on the moon. Find a recipe that uses only snap peas and five other ingredients. Explain quantum physics. Paint a dinosaur with a briefcase in the style of Van Gogh. When researchers started developing artificial intelligence in the 1950s, they could've never predicted what humans would ask of it. But beyond AI's entertaining or day-to-day search functions, the technology has the power to transform our lives. Oren Etzioni, Founding CEO of the Allen Institute for AI, recently joined Round to discuss how artificial intelligence evolved and what he sees as AI's most promising — and concerning — consequences. Here are three key takeaways from the conversation.
Although artificial intelligence has been an active field of research for decades, before 2018, AI had narrow practical applications — like predicting airfare fluctuations or recognizing spam emails — for two reasons. First, they required humans to label data sets, which was expensive and reduced enormous data sets to a fraction of their original size. Second, early AI models needed narrowly defined targets, and to change the output, you had to retrain the system. (For example, an algorithm trained to recognize images of breast cancer could not identify tumors of any other kind of cancer) . In 2018, researchers began training new models on massive amounts of unlabeled data (like the billions of sentences written in English on the internet) without human supervision, and the predictive power of AI exploded. Soon, tools like ChatGPT utilized AI predictions to write original stories, create highly stylized selfies, generate code, and more.
Earlier this month, ChatGPT falsely accused a George Washington University law professor of sexual harassment and forged a convincing Washington Post article as evidence. How did this happen, and why? Egregious errors like this one, dubbed "hallucinations" by researchers, are poorly understood. Etzioni predicts we'll develop tools to curb hallucinations but still fears AI's' potential to spread misinformation through forged articles and convincing deepfakes. AI also has a bias problem. For instance, when someone asked ChatGPT, "Who are the ten greatest philosophers?" it gave ten white males. When asked to broaden its answer, it did but then reverted to its original response when asked the original question again. Plus, maintaining large-language models (LLMs) presents technical challenges, including the high cost of running, training, and querying AI, not to mention the huge amounts of electricity required to power it.
Some have compared the current AI explosion to the emergence of the early internet. Like the internet, AI may be poised to radically transform our world, revolutionizing fields as diverse as education, medicine, and transportation. And even AI hallucinations may have a silver lining. Etzioni says the other side of forgery is creativity. People can use AI to paint like Picasso or make music without picking up an instrument. The musician Grimes, for instance, recently announced on Twitter that she'd split her royalties 50/50 on any AI-generated song that uses her voice. Etzioni also highlights AI's potential in healthcare as researchers begin training their models on biological data. Moderna allegedly used AI to help develop its COVID-19 vaccine, and in the future, AI could prove a powerful tool for diagnosing and curing intractable illnesses. We've yet to see all that AI can do, but as Etzioni assures us, we're on the brink of something enormous.
Jad Esber, CEO of koodos labs and Harvard Affiliate, discusses the potential of web3 and how it can create value for users, creators, and enterprises.
Sonny Patel, VP of Product Engineering & Conversational Intelligence at LivePerson, discusses the latest advancements in chatbots.
What happens to humans when computers can make art or write books? Should we be worried? Greg Hochmuth explains the recent explosion in AI creativity.