- AI thinks very differently from us, raising trust issues.
- It's unclear whether AI is inspired by use or just copies us.
- AI will change the job market, for better or worse.
Can we trust artificial intelligence? Will it take our jobs? Can it be creative? Those were some of the topics that came up last month when the Portugal Economic Forum met in New York City and I moderated a panel on “The Future of AI.” (Disclosure: I received food but no other payment.) I had the pleasure of speaking with four Portuguese scientists and entrepreneurs:
- Manuela Veloso, Head of AI Research at JPMorgan Chase, former head of the Machine Learning Department at Carnegie Mellon University
- Daniela Braga, Founder & CEO of Defined.ai, which provides training data and models for industry clients
- Vasco Pedro, CEO & Co-Founder of Unbabel, which centralizes language translation for companies
- Carlos Costa, Co-Founder & Chairman of Armis Group, which provides consulting on data analytics, big data, visualization, and business intelligence
Below is a transcript edited for brevity and clarity.
Veloso: Robots that we had at Carnegie Mellon would move down the corridors by themselves and transport items from one place to another or guide visitors. They would arrive to my office, and I wanted to know: What did you do? Similarly, I’ve done a lot of robot soccer, and many times you would see these robots not passing the ball to each other, or missing a shot or succeeding, and you’d like to know: How in the world did this happen?
In the financial world, again, we developed methods to explain what these machines do, for three types of purposes. One, just trying to debug what's happening. Then there is a problem of the regulators that actually want to know: Why did you make that decision? And then there is also the problem of responding to the users. If the user cannot have a credit card, you need to have an explanation if the user questions why.
And one final thought. The financial domain is a little bit more complicated, because you cannot provide explanations that are what we call not actionable explanations. You cannot tell someone: If you just had fewer kids, or if your age were different, you’d get the credit.
Pedro: The number of kids could be actionable.
Veloso: To decrease the number, I hope not. [Audience laughs.] And one final thought, sorry. Waze moved from just giving us a route to showing the greens, the yellows, and the reds. So you have a little bit more trust that the route they are recommending is probably better than something else. In contrast, Netflix sends these recommendations about what you should watch without any explanation. So I’m very upset with Netflix. [Audience laughs.]
Hutson: Explanation is especially important when dealing with other kinds of minds. Vasco, how different are human and machine cognition?
Pedro: First I just want to say that it’s quite interesting that we want to know the same thing about soccer players whether robotic or not. Why is that guy hogging the ball? [Audience laughs.]
We understand very little about our own cognition, and so to the extent that we tried to replicate it in AI, it’s usually more on inputs and outputs, whether it’s behaving similarly to what humans do. We’ve seen the news recently, researchers at Google saying some of the neural networks are developing consciousness. Well, the reality is, we don’t know what consciousness is. So when will you know if a machine has it or not?
Hutson: Especially mysterious is the black box of machine learning. Daniela, could you explain the rise of ML within AI, and why data is the new oil?
Braga: Instead of programming rule-based programs, doing if-then processing, we’re throwing data at an artificial brain. And this is where explainability comes in, because then you don’t know what’s going on anymore. Data is the new oil, because it represents about a third of the investment in AI. When I was at Microsoft, we would spend at least a third of our AI budget on collecting and structuring data. This is also why I started Defined.ai.
Pedro: I just wanted to say that one of the most interesting developments I’ve seen over the last few years in AI is the democratization of the ability to do things that are relevant. The tools are becoming much more operational. This summer, I was actually going through a deep-learning book with my 14-year-old. This is how simple it’s starting to get.
Hutson: Carlos, you’ve worked on finance, sports, security. What’s been your favorite application of AI?
Costa: For our first big project in AI, we collected five years of data on traffic to predict or prevent accidents or traffic jams. We now have simulations to train autonomous vehicles.
Veloso: Connecting with what Carlos was saying, synthetic data enables us to do all these counterfactuals: What if the world were like this? Then we can become creative. I mean, who wants to learn that there are traffic jams? No, we want to solve the problem of traffic jams. AlphaGo learned to play Go by looking at games of Go that have been recorded, but then it played itself to say: What if I do this?
Pedro: There are a lot of researchers who believe that the ability to model the future is an essential component of human intelligence. Without having a mental model of what can be, you can’t be creative.
Veloso: Imagination is simulation.
Hutson: Traffic is a life-or-death situation. How do we get people to trust AI, but not too much?
Pedro: We inevitably will use things that give us advantages. As soon as we get correct predictions from AI two or three times, we stop really questioning if it’s correct.
Veloso: To build trust, what we should have are AI systems that say “I don’t know how to do this” when it has low confidence.
Hutson: How do you see AI changing the job market?
Pedro: We’re seeing the incredible explosion of AI-driven creative tools, for writing, for images, and for video. That creates concerns. Does it mean that designers are going to be out of a job? In some ways, yes. But it also means that many more people are able to produce creative content that is quite spectacular. If you don’t need a $100-million budget to create an amazing movie, and you can do it from your computer, then there’s going to be 100,000 more movies made. Or instead of having a team of 10 designers, you might have one designer who uses AI tools to produce 10 times more work.
Braga: It should be regulated. Companies are starting to scrape the web for creative artists’ work to train their models. That’s why we exist as Defined.ai. Scraping the web goes against all the data privacy, all the data copyrights, all the problems that people don’t even know.
Pedro: The guy who won an artistic contest with an AI-generated image, he actually spent months refining that image. AI gave him the ability to express his creativity.
Braga: I understand, but we should not steal content. We need to regulate that.
Pedro: We can make a bet that in the next five years, there’s going to be a movie on Netflix or in the movie theaters that’s going to be solely produced by one person with the help of AI tools.
Braga: That’s fine. As long as it doesn’t take the style of author A, B, or C.
Veloso: I think both of you are right. As soon as an artist exposes one of their creations, immediately its open for someone to get inspired and do something similar. How can a Picasso come up with their own thing by not being inspired by anything else? In music, the same thing. There are people who play like Mozart. You can have all the regulations you want, but the fact is that humans are like this.
Braga: You gotta pay the author.
Veloso: It’s not stealing, it’s being inspired.
Costa: During the last two years, we have learned some lessons. Covid killed a lot of jobs, and people found other types of jobs.
Pedro: Technology so far has always created higher GDP, and when you correlate GDP with number of jobs, there’s always been more jobs with more GDP. So if AI increases GDP dramatically, there’s no reason to think that suddenly we’re all gonna be out of a job. I think some people will have to change, but I have no doubt that the creation of value is going to benefit everyone’s life.
Veloso: I’m part of a study at the National Academy of Engineering to study AI and its impact on the workforce. Think about some company in which the job of like 10% of the people is to read documents online. Then think, could we have an AI system that at least whispers to them what these documents say? Some of those people will eventually move on to do other things.
Now unfortunately, the type of jobs that you are not going to have AI in the near future do is like janitors. We think about autonomous cars, but we don’t think about all these manipulation tasks we do.
I wish that some AI system would answer all my emails, that the AI would help free me to use my brain for other things.
Pedro: Think about an orchestra. Right now, a lot of humans are very specialized instruments. I see that more humans are going to become conductors, where AIs are going to become specific instruments and humans can use that to produce more rich music.