“Everybody is talking about AI and everybody is confused about how intelligent or not today’s AI systems are,” says SFI Professor Melanie Mitchell. “This whole notion of what, exactly, is intelligence is extremely fraught.”
Views on intelligence have changed over the years. Computing systems were once deemed intelligent based on benchmarks such as skills in games such as chess, Go, or the ability to use language. However, it has also become increasingly clear that intelligence is not associated with a particular skill or ability. What, then, is intelligence?
On August 19–23, Mitchell and SFI External Professor John Krakauer (Johns Hopkins University) led a working group on “The Nature of Intelligence.” It was the first in a series of six meetings to be held over the next three years. Scholars from diverse fields — neuroscience, psychology, linguistics, philosophy, and AI — were invited to investigate the broad notion of intelligence, whether in machines or biological systems.
“We are getting good at human cognitive science and at playing God with non-human animal models. We are in this unbelievable position of dissecting, to a very great degree of invasiveness, so many forms of intelligence,” says Krakauer. “Surely we can start coming up with some sort of general set of principles, because we have so many sources of information about so many types of intelligent behavior. It seems like a very pregnant moment.”
The August working group delved into questions about different dichotomies in intelligence: such as conscious, deliberative thinking based on predictive models versus non-deliberative, reflexive behavior. Working group participants hashed out distinctions between language and thought, comprehension and competence, and “world models” and automatic behavior.
“We brought out a lot of key areas where people disagree and where some key ideas really need to be nailed down,” says Mitchell.
Participants also talked about how large language models could be characterized, what they are capable of, and how their intelligence relates to that of humans. They explored questions about the evaluation techniques to determine the intelligence of such models and what would be needed to ensure that such evaluations are scientifically sound.
This first meeting laid the groundwork for the remaining five working groups. The organizers expect future meetings to further examine the key concepts and ideas raised in the August meeting, and plan to invite additional participants to broaden perspectives.
“SFI should be a beacon for investigating the nature of intelligence across animals, humans, and machines, and create an environment to go after it quite broadly, embracing disagreements and controversy,” says Krakauer.