Maturing technology offers plenty of both promise and peril, say industry experts
Recommended Videos
“A lot of jobs are going to be replaced,” he said. “We don’t know at present whether it’s going to lead to massive joblessness, but my guess is it probably will. Mundane intellectual labour is just going to be replaced by AI.”
“If it does lead to massive job loss, we’re going to need something like universal basic income,” he said.
Jacob Steinhardt, an assistant professor of electrical engineering and computer science at UC Berkeley who also spoke at the Hinton event, said he is hopeful everyone will join in the effort to discuss AI and determine how it should best be used, as well as what safeguards are needed.
“Given the kind of stakes and impact of AI in society,” he said, “we need all of society to be thinking ahead and planning for or what’s going to come.”
That includes anticipating the ways in which AI models may learn bad habits. One such problem is the concept of rewarding hacking, where a program (called a model in industry terms) will attempt to achieve a desired outcome with creative strategies that aren’t intended, thereby essentially gaming the system.
Steinhardt used an example of a model tasked with optimizing highway traffic patterns that eventually decided that closing an on-ramp was the best way to achieve the desired goal even though that prevented cars from entering the highway altogether.
As models get larger, they may also create an echo chamber for the user, offering deceptive answers instead of admitting they don’t know, or engage in sandbagging, where a model intentionally does not perform up to its capabilities. In one form of sandbagging, a model may give less accurate answers to users it thinks are less educated.
“How big of a problem is it if chatbots are not always honest? I’d say it’s actually a pretty big problem,” Steinhardt said.
One amusing example he gave was of an insistent Bing chatbot gaslighting a user into thinking it was a different year, but he pointed to how such behaviour could be used on a large scale for harmful purposes.
“If someone were to try to train for persuasiveness, perhaps an unscrupulous company or a government that cared about persuading its citizens, there’s a lot of things you could do,” he said. “If you had a large user base, you could actually fine-tune strategies on thousands or millions of users in parallel. There’s a lot of data you could collect just to practice being persuasive.”
“It’s getting more and more difficult to tell the difference between AI-generated videos or voices or images and real ones,” Hinton said.
Still, Hinton offered some optimism in the ways larger AI models can help people.
Enabling people to learn more is another positive effect.
“We’re all going to have our own personal tutors who understand what it is we misunderstand and explain things to us so we can understand them,” Hinton said. “Research shows that if you have a personal tutor, you learn stuff twice as fast as you do in a classroom, and with an AI personal tutor, it’ll probably be three times as fast.”
Steinhardt offered solutions as well, such as creating AI models to monitor other AI models and fostering an informed and engaged society that provides input on AI uses and limits.
“I think there needs to be an entire public ecosystem for understanding AI models,” he said. “By default, this is just going to be a conversation that happens at companies, but I think that can’t be the case. The public needs to participate in this conversation. These systems are really complicated, and we can’t just rely on a few scientists or a few engineers to try to figure this out.”
From mass joblessness to cyber crime to biological attacks, there’s plenty to be concerned about with AI, but Steinhardt positioned himself as a “worried optimist,” and his vision of the future allows for varied outcomes.
He speculated that there’s about a 50 per cent chance AI leads to radical prosperity and a 40 per cent chance of a future with a kind of complicated tension between success and chaos. The alternative is less ideal.
“I’d say there’s maybe a 10 per cent chance that this kind of all blows up and we go extinct or something similar happens,” he said.