Dean's Distinguished Speaker Melanie Mitchell Maps the Development of Artificial Intelligence

"What exactly is this thing we call 'intelligence,' and how is it different in, say, humans versus machines?"  

This was the first of many big questions posed by College of Engineering Dean's Distinguished Speaker Melanie Mitchell when she visited the University of California, Davis, on March 7. Mitchell, a professor at the Sante Fe Institute and pioneer in the field of artificial intelligence, presented "The Past, Present and Uncertain Futures of AI" to nearly 150 faculty, staff and students. 

"For me, it's really been sort of a journey in understanding what it is to be human," she said. 

And what a journey it has been. Since its inception at a conference held at Dartmouth College in 1956, AI has been regarded with peaks and valleys of optimism and public opinion. A notable dip came at the end of the 1980s when Mitchell finished her Ph.D. and was discouraged from mentioning artificial intelligence in job applications. Currently, in the era of generative AI, Mitchell notes that "optimism has gone through the roof." 

While diving into what she described as "the astounding, hopeful, terrifying and confusing present," Mitchell demonstrated several remarkable exchanges with ChatGPT.    

"It's a poet, it's a mathematician, it's a translator, it can be an artist," Mitchell said about the technology. "All of us, including people who have worked in AI for decades, are very, very surprised by this." 

Mitchell described how chatbots work, including the pre-training, supervised fine-tuning and reinforcement learning from human feedback required to make it "nice" for humans to engage with. 

"It's an amazingly complex thing," she said of ChatGPT 4, the latest version of OpenAI's language model system launched in March 2023. "And it's so complex that it's hard to know exactly what it actually has learned." 

With the recent emergent abilities of large language models, ChatGPT has passed business school, bar and medical licensing exams, leading some to believe that AI is approaching "general intelligence." However, Mitchell's own research in AI has found its abstraction and reasoning to be far less than humans. For instance, when asked to solve a simple shapes-based puzzle, the chatbot failed the majority of the time, whereas humans succeeded nearly 100% of the time. 

To conclude her presentation, Mitchell described her hopes and fears and raised several more questions about the future of AI.  

"How can AI learn to better understand our world, our values, our intentions? And, can we develop the scientific tools to understand AI?" asked Mitchell. "The future is not inevitable. It is up to us to create it."