AI RISING: Risk vs Reward – The Hinton Lectures™ - Livestream Video - Day 1

Thank you. Hi. As Sonia said, the Global Risk Institute decided to fund a series of lectures they called The Hinton Lectures, which was nice of them, but I'm not giving them. They're about AI safety. AI safety has three main categories, I think: there's bad actors doing bad things with AI; there's accidental disasters caused by AI not doing what you expected it to do, that's the alignment problem; and then there's the existential threat that AI will get super intelligent and take over from us. So if you look at the bad actor threats, there's things like cyber crime. Phishing attacks increased 1200% last year, largely due to large language models being very good at them. There's biological threats. It's fairly cheap now to design things like COVID. There's joblessness. AI is going to cause big increases in productivity, and we don't know yet whether that's going to cause massive joblessness. It might, it might not. If it does, it's a serious problem we have to deal with, and it's a serious problem that could make society worse by increasing the gap between the rich and the poor. The danger is that this increase in productivity will all go to the rich. Then there's fakes. It's getting more and more difficult to tell the difference between AI-generated videos or voices or images and real ones. For a while I thought we may be able to label things as generated by AI. I think it's more plausible now to be able to recognize that things are real by taking a code in them and going to some website and seeing the same things on that website. I think that's what the White House believes too, that recognizing they're real is a much easier problem than recognizing that they're fake. There's lethal autonomous weapons. If you look at the regulations people are introducing, like the European regulations have a clause in them that says none of this applies to military uses of AI. Costa Rica, however, actually is against using lethal autonomous weapons. It's actually introducing legislation against them, so good on Costa Rica. Then there's discrimination and bias. I'm an old white male, so I don't experience it the same way other people do. I think that's one of the problems for which there's some sort of a solution, which is don't aim to make things completely undiscriminating or unbiased, just aim to make them better than what they replace. I think that's doable because it's much easier to measure bias in an AI system than in a person. If you start measuring bias in a person, you get what's called the Volkswagen effect, which is they start behaving differently because they know you're measuring them. With AI systems, you can freeze the weights and really find out how biased it is and make it a bit less biased. So the Global Risk Institute came to me and said we'd like to create something. Initially they wanted to create a prize. I thought they'd be better to create a lecture series, and with the lecture series we can find the sort of best people in the world and get them to come to Toronto and give several lectures. That's what the idea of the Hinton Lectures is. We set up a committee of very good people in AI safety, and Jacob Steinhardt was the first choice of quite a large number of them. So we'd like to welcome him back to Toronto. He actually applied as a graduate student here and we made him a very nice offer and he didn't come. He foolishly went to Stanford and got a PhD with Percy Liang. Then he became an assistant professor at Berkeley. He's done work on understanding AI systems, he's done work on aligning AI systems. A little while ago he had some AI benchmarks which were downloaded 100 million times. He's designed AI systems to explain complicated systems, he's designed AI systems that'll give well-calibrated forecasts. He worked at OpenAI, Open Philanthropy. He's got lots of awards, actually a surprising number for someone so young. And very recently, last Wednesday, it became public that he's just set up a nonprofit working on AI safety called Translu. So without further ado, let me welcome Jacob.