[0:00:00] 2023 was the year that artificial intelligence burst into public consciousness, stoking the imaginations of many and the fears of many others. In 2024, Goldman Sachs says generative AI will move from an excitement phase to a deployment phase, and in the next 10 years, could raise global GDP by as much as 7%.
[0:00:25] What gets you interested in artificial intelligence? Well, I've always been obsessed with the idea that tools and invention
[0:00:51] are going to make us more capable as a species. And that's been the history of human civilization. We've invented things in order to make our lives easier and happier and healthier.
[0:01:04] From a kind of social impact perspective, that's the grand offering of science. If we look back over the centuries, our lives have got radically, radically better. And that's real cause to be optimistic. This engine of discovery and creation is fundamental to who we are as a species. Our curiosity drives us to create and do more with less. And that's actually what drew me to AI in the first place.
[0:01:32] Most people will understand, get the AI to do something specific. Let's say look at a legal document and find all the cases that relate to, or look at an x-ray and figure out, compared to all the other x-rays, does this one have cancerous cells? Help us understand generative AI, which is the ability...
[0:01:54] Think about it as two sides of the same AI coin. For the last decade, we've been focused on classification. How can we train deep neural networks to identify patterns in data?
[0:02:17] And it's done that so well that it can transcribe speech, for example, or it can understand the content of images. And that's been the classification revolution. The flip side of that is if the AI now understands the difference between cats and dogs and zebras in an image, you can then say, generate me a new example of a zebra that's never been seen with the color purple and with bright...
[0:02:43] Blue Wings, for example, because it understands those concepts well enough from having seen the historical patterns. Now, the third phase of AI that's coming is interactive AI. Now that you can have an AI generate a new example, it can present that to a human for feedback and interact in order to learn from the human and be useful to you, carry out new tasks. In time, it will learn to use tools. So it will actually be able to make phone calls.
[0:03:11] Let me just get an example out so people understand. Correct me if I'm wrong, but I think what you're saying is you should be able to have an AI that you say to your phone, can you do the following for me? Two weeks from now, I want to take a trip to London. Can you get me a ticket, book me a hotel? I want to go to this restaurant, make me the reservation. And I also need these three books. Can you order them?
[0:03:40] And the AI will then take care of it all, right? Exactly. Think of AIs in the future as personal assistants. Everybody is going to have a conversational interface, which not just represents you and is there to help you and support you, but can also teach you, is actually going to help you day to day to be better at your job.
[0:04:02] to make important life decisions when you're thinking about whether to relocate to a different city or whether it might be time to put your elderly parents into a care home or whether to go ahead and have a serious operation that you're thinking about. So that gets to a very interesting point, which is emotional intelligence. We've often thought that the AI will do the analytic part, but human beings have this special thing, emotional intelligence, EQ, social skills.
[0:04:32] The AI can have all that as well. That's exactly right. I mean, we have developed at my new company, Inflection, an AI called PIE, which stands for personal intelligence. And we've specifically conditioned it to be good at emotional intelligence. It's a great listener. It's very even-handed. It presents both sides of an argument. It asks you great questions. It tries to remember what you've said in the past. It'll often paraphrase what you've said in order to make you feel heard.
[0:05:02] Sort of elements of empathy are actually quite learnable by the AI and will be incredibly valuable to millions and millions of people to have the comfort and support of a friend, a companion, an emotional intelligence in your pocket. I mean, I think people underrate how significant that is in the world. Not all of us have a supportive parent or...
[0:05:26] How do we stay in charge? That's the key question. And I think in the limit, it's definitely possible that...
[0:05:42] We may not be in charge, and we have to confront that reality that isn't desirable. It's not what we want. We don't want AIs that have autonomy in our world. They shouldn't have rights, as some people have proposed. We don't want new digital citizens hanging around in our everyday society. Just so that people understand technically, is it possible to say, is it as simple as saying, look, at the end of the day, you can always have a kill switch.
[0:06:12] You can pull the plug. Or is it more complicated than that? Yes, it's going to be possible to do that, I believe. If we focus on it as a research community, as a group of governing actors, as creators of the technology in the open source and in the private companies, we have to make that a priority. We have overcome huge scientific and engineering feats to contain nuclear power.
[0:06:37] Get the best out of aviation and not have planes fall out of the sky all the time. To regulate cars and make sure that everybody is trained to drive a car and that there are airbags and windscreens and vehicle emissions and traffic lights. These are important moments in the history of the development of our technology where we've really paid attention. And this is another one of those moments. The good news is that over the last decade, as the models have got bigger,
[0:07:07] They've actually got easier to control. There are fewer hallucinations than they have ever been. There are fewer biases than they've ever been. And with each order of magnitude of computation that we add to these training runs, what we see is that they make fewer mistakes, they're easier to control, and they take instruction more closely. And so the real question becomes, whose instruction and what instruction?
[0:07:33] What is the behavior policy that we're asking these AIs to adhere to? And who is auditing and verifying that the AIs are behaving within the boundary constraints that have been set for it? Tell me, when you look at the future of AI, do you worry about something that Henry Kissinger and Eric Schmidt have written a lot about, which is that human beings will really be in a position of not understanding the world anymore?
[0:08:03] Before the Enlightenment, we didn't understand the world, and we would attribute anything we didn't understand to divine forces. The sun rose because the sun god moved his chariot across the sky. We will be in a situation where the AI will tell us, this is the answer, this person has cancer. But we won't be able to understand why, because the level of computation and analysis will be so much greater.
[0:08:33] What kind of position does that place human beings in? Where we're almost returning to our pre-enlightenment self, where in that time we looked up to God and now we look up to this black box called AI. But in both cases, we're believing, not actually understanding. It's a great point and it is a real risk. And we can't rest on belief. We have to rest on reason. You know, belief is a dangerous path to...
[0:09:01] Being able to justify actions on the basis of things that we can't interpret. At the same time, we are all behaviorists today. We observe outputs. I don't really understand what's going on inside your mind, for example. I rely on the things that you say and the things that you do to verify that you can repeatedly act consistently with your own behavior policy, your own set of values. And that's actually how we form trust.
[0:09:29] What is the role of AI in society today?
[0:09:55] But at the same time, we have to be realistic about the benefits. Take healthcare, for example. These models are performing an incredible role in being able to detect disease at human level performance across the board on radiology scans, for chest x-rays, for mammograms, for acute kidney injury, for sepsis, for glaucoma and blinding eye diseases.
[0:10:21] And the cost of producing a world class diagnosis and a proposed treatment pathway is about to plummet to zero marginal cost. It is going to be essentially free in the next decade to provide an absolutely world class diagnosis. That's going to transform health care for billions and billions of people. And we're going to see the same trajectory.
[0:10:45] In education, we'll see the same trajectory in the development of food and synthetic materials. And we'll see the same trajectory for decision making and entrepreneurship across the board. So the benefits are really incredible. The challenge is how can we mitigate those downsides?