There are areas where machine cognition still lags far behind that of humans. AI systems’ ability to plan is currently “at a child’s level,” Bengio says, although he notes that frontier models are making rapid progress in this area. Because they are trained primarily on text and images, their spatial reasoning skills are also poor, he says.
A big question in contemporary AI research is how many further models can be pushed toward human or superhuman intelligence. While companies have made tremendous progress by packing enormous amounts of computing power into their models, it’s not clear that increasingly powerful computers will result in increasingly smarter machines. At some point something else may be needed.
Mitchell notes that while babies learn by interacting with the world around them, AI systems are trained passively by receiving vast amounts of information. This may help explain why chatbots are prone to lying: in the absence of real-world feedback, they often have difficulty distinguishing whether information is true or false.
Bengio says machine intelligence will continue to develop unevenly, rather than crossing a threshold into a world of conscious machines as we see in science fiction films.
“We shouldn’t think about it: have we passed the threshold of AGI?” says Bengio. (AGI, or artificial general intelligence, often refers to machines with human-like intelligence.) “Because of the erratic way intelligence is evolving in AI, there may never be such a moment.”
