Lynn Thoman

Ep 311: Alphabet (Google) chairman John Hennessy on The Future of AI, Human Intelligence, and Leadership

This transcript was auto-generated. Please forgive any errors.

Lynn Thoman: Every generation believes it's living through the biggest tech revolution in history. Usually it's wrong, but every once in a while a technology comes along that changes almost everything. Electricity, the microprocessor, the Internet.

Is artificial intelligence the next technology on that list, or are we dramatically overestimating its impact? And if this is really a once-in-a-generation moment, what are most of us still missing?

Hi everyone, I'm Lynn Thoman, and this is 3 Takeaways. On 3 Takeaways, I talk with some of the world's best thinkers, business leaders, writers, politicians, newsmakers, and scientists.

Each episode ends with three key takeaways to help us understand the world, and maybe even ourselves a little better.

Lynn Thoman:Today, I’m delighted to be with John Hennessy.

John is one of the world's leading computer scientists. He co-invented RISC architecture, which helped lay the foundation for modern computing. He received the Turing Award, often called the Nobel Prize of Computing.

He also served as president of Stanford University, and today he is chairman of Alphabet, Google's parent company. Few people have had a front-row seat to as many tech revolutions or helped shaped so many of them.

John, it's great to be with you again. Welcome to 3 Takeaways.

John Hennessy: Thank you, Lynn. Delighted to be here.

Lynn Thoman: It is my pleasure.

Let's start with why this moment is different. You've had a front-row seat to almost every major computing revolution of the last 50 years.

Is artificial intelligence the biggest one?

John Hennessy: I would say it is, for one particular reason. It changes everything. The Internet certainly changed the way we do a lot of things, the way we interact with people, the way we shop, the way we do things.

But it didn't change some more fundamental things about how we operate. Same with the microprocessor, the personal computer, it changed various aspects. But artificial intelligence is going to change the way we do science, the way we interact with people, the way we get things done, the way we get customer service, for example.

So I think because of its ability to replace function or augment human capability, that makes it different.

Lynn Thoman: How will it change each of those things?

John Hennessy: I think you can already see this in science. I mean, you saw this with the AlphaFold breakthrough that DeepMind did, where they solved a long-standing problem we call a protein folding problem, which is to determine the 3D structure of a protein, which determines how the protein actually interacts in our body in various ways. That problem has been there for 40 or 50 years.

They made this big quantum leap in terms of their ability to solve it. They increased the size of what we know about protein structures by an order of magnitude using this technology. So I think that's one example we see it happening in science.

We see it happening with coding. People who are software engineers all use AI as part of the tool for doing their coding.

Lynn Thoman: If we were looking back 10 years from now, what would have to happen for people to say AI has truly changed the world?

John Hennessy: I think we've already achieved sort of the first hurdle in what we think of as intelligent behavior, which is a long-standing test invented by Alan Turing called the Turing test. And what the Turing test says is, if you're carrying on a conversation across the internet, let's say, and you can't tell if you're talking to a computer or a person, you've reached that level of intelligence. That was Turing's original test.

This is done. You cannot tell that you're talking to an LLM versus a real person right now. But they still don't have broad intelligence across a variety of fields.

There's a great benchmark called humanity's last exam, and it's a collection of 2,500 hard problems. Hard meaning they're college-level seniors for people who are majoring in that field, for example. They're doable problems, but they require a level of sophistication in education.

Right now, the best systems in the world achieve about a 60% grade on that. They solve 60% of the problems. So, as I remind everybody, as a long-term professor, 60% is a failing grade.

So we still have a ways to go in terms of broad reasoning intelligence, the ability to reason, which is what humans can do really well.

Lynn Thoman: Everyone talks about artificial general intelligence as if it's a destination, but what is the actual bar? What would have to be demonstrably true for you to say we've crossed the line?

John Hennessy: Yeah. The first observation you make is everybody talks about doing it, but there is no widely accepted definition of what it means. Other than you can reason across a wide range of problems, you can compete, let's say, with people who would regard themselves as professionals or well-educated in a field.

We're getting there, but the thing to remember about the current state of the LLMs [large language models], a lot of what we see is intelligent, like answering questions. What is the capital of France? It's Paris.

Well, how does it know the capital of France is Paris? Because it's read every single thing, document on the Internet, and it's read that thousands of times, and it's basically memorized that connection between the two. And it's parroting back that connection.

So lots of intelligent behavior is memorization. And the real challenge is to get to reasoning, which really makes us different. And when we get systems that can really reason through complex problems, I think then we'll know it's AGI.

I'll know it when I see it.

Lynn Thoman: So suppose we do cross that line of reasoning to artificial general intelligence. What becomes possible that simply isn't possible today?

John Hennessy: Well, I think we could automate a lot of functions that are very hard to do. And it'll get used in different ways in different parts of the world. I think if you look at developed countries like the US, it's quite likely that AI will be an augmentation of what people can do.

So it'll be there to help a physician sort through a diagnosis of a difficult case. It'll be there to help students and help a teacher who's struggling with students at various levels of capability in a classroom. It'll be there to support legal systems and help get better access to legal services.

But it'll also do something very different in other parts of the world. In so much of the world in the global south, there's a shortage of medical capability. There's a shortage of teachers.

There's a shortage of lots of things that require more sophisticated education. This technology could really help them advance at a rate which currently they're just, they don't have the human capital to do it. So that could really be a real way to raise the standard of living and the domestic product around the world.

Lynn Thoman: As chairman of Alphabet, the parent company of Google, you're in a very unique position. What's the single biggest idea about artificial intelligence that even well-informed people still don't fully understand?

John Hennessy: The most important thing to understand is that these systems learn in a very different way than people learn. They learn by taking massive amounts of data. If, for example, to have a child in school capable of reading and writing, we had to first show them thousands and thousands and thousands of documents before they were able to read or write.

We would never get to that. Humans have an ability to do intelligent things that is quite remarkable. It's astonishing.

You know, a human brain consumes about 30 watts of power. A data center consumes thousands of watts of power. So there's an enormous difference in terms of our ability to do things efficiently.

And if we ever want to achieve something that really comes close to human capability, we're going to have to get more inspiration from how the brain is organized and figure out what can we take advantage of to make something that's much more efficient than the current designs.

Lynn Thoman: There's a growing debate about whether AI will concentrate power in a handful of companies or democratize it by giving billions of people extraordinary new capabilities. Which future do you think we're moving toward?

John Hennessy: I think you're going to see AI propagate down. It may be that some parts of the technology stack are dominated by a few players. That's been true in the technology business for quite some time, right?

I mean, in the PC era, it was dominated by Microsoft and Intel, for example. That's been true for a while. But the real power of AI comes from applying it in the context of some particular area or problem you're trying to solve.

And that will happen in a much more democratized fashion with lots of people building pieces of that stack.

Lynn Thoman: And also the fact that some companies have proprietary large language models, whereas others have what are called open weight models that are available to anybody to download or to revise or use.

John Hennessy: Correct. So right now there's an ongoing debate about open source versus closed weight models. And most of the companies have both, even though their leading edge model is probably a closed weight model and they have other open weight models.

There are various complexities in this. One of those problems we're struggling with is if you train the model and you try to put what we call guardrails, things that prevent misuse of the model on, it's virtually impossible in an open weight model because people have access to the weights. They can strip off the weights that are the guardrails.

Even in closed weight models, it can be hard to get guardrails that are really secure in that. So that's a debate that's very much going on and how those models get used and how they get replicated. I think one of the struggles we have is that you'd like to prevent all negative uses.

The problem is it's a software technology, so it's flexible. It's inherently flexible. That's what gives it its power.

But that means that it's easier for somebody to try to misuse the model.

Lynn Thoman: Many people assume that the biggest danger from AI is science fiction, a rogue super intelligence. From your vantage point, what's the real world risk that worries you the most?

John Hennessy: The one that worries me probably the most is use of AI in weapons systems that would be autonomous. So there's no human in the loop anymore. You're not just identifying targets and various other things, which I can imagine an AI system would be useful to do, but you're actually letting the AI system decide to fire a missile or do something else that could lead to a nightmarish scenario because humans would no longer be involved in war.

You could make it in a way that could be extremely dangerous. We've already seen a bit of this in the drone battle that's going on in parts of Ukraine and Russia, for example, or in the Middle East. I think figuring out how we're going to prevent that, just as we did with, say, a poison gas and a ban on biological weapons in an earlier time, that has to be negotiated at a multilevel, global level.

Lynn Thoman: That certainly is a nightmare scenario. On the flip side, what's the opportunity that excites you the most?

John Hennessy: One that particularly excites me as a lifelong academic and researcher is that we could get dramatically improved rates of scientific discovery to work on really important problems. Imagine unleashing this technology on finding a solution for climate change, for example, figuring out how to do nuclear fusion and really make it work so we'd have a carbon-free energy source, figuring out how to find solutions to Alzheimer's or Parkinson's. Those are things which really excite me because I think the positive impact on human existence could be enormous.

Lynn Thoman: Those would be phenomenal. And what you're really talking about is having people creatively ask AI questions to advance research.

John Hennessy: And what we call these agentic systems, where AI helps explore the design space and consider a variety of different alternatives using intelligent models to sort out which of those alternatives are most attractive. This doesn't eliminate humans from the research, but it empowers them in a different way.

Lynn Thoman: So using AI to explore new ideas, then. How should people use AI?

John Hennessy: I'm a big believer that the primary focus of AI is augmenting human capability, not replacing it. So using it to polish a document that you're writing is a great idea. Using it to help you think through various ways of doing something or composing a communication.

Using it to simply replace doing the human part of it is probably not a good idea. You may not be happy with the results. So that's something you want to think about.

How do you keep in the loop on anything you're creating? You know, we say about people who are using AI in research is you're responsible for the results that get published, not the AI system. You can't blame if there's something wrong in that scientific paper.

Lynn Thoman: So you're really talking about using AI to leverage humans.

John Hennessy: Leverage, augmentation, leverage capability. Yes.

Lynn Thoman: Billions of people use Google search every day to decide what information they trust. As AI shifts us from searching for information to simply asking for answers, what happens to society if fewer people know where those answers come from?

John Hennessy: If you use Google search today and you're in AI mode, you'll get a citation for where it comes from. So I use those citations to follow up. If I want to learn more, I can click on that and then I can get more details about where it came from.

We're going to have to be careful with ensuring information that comes from an authoritative source, because one of the things that will happen as more and more of the things documents on the Internet were created by AI, it becomes easier to have a hallucination propagate because one thing read it and it replicated something that was wrong and somebody else replicated that. So we've got to figure out how we distinguish between informations where we know that the facts are accurate and situations where we might not be cognizant of that.

Lynn Thoman: Yeah, that's so important.

AI tutors may soon become better than human teachers for many subjects. If that happens, what is a university actually for 20 years from now?

John Hennessy: There is some interesting research on exactly this area. And what it's demonstrated is that students working with AI tutors alone do well. The tutor improves the ability of the student to get their work done as opposed to a student working without a tutor.

But the best combination is student, teacher and tutor. And the tutor, then the AI tutor, can take guidance from the teacher as well as understanding from where the student's gap is and combine them. And I think we're going to see more of that model where people are not pulled out of the loop, but where the AI tutor is an assistant to either a human teacher or a human tutor for that matter.

Lynn Thoman: And how will that change universities?

John Hennessy: I think universities are going to have to change right now. I mean, we're already thinking about this. I'm an engineer, right?

I'm a computer scientist by background in computing and engineering. And the science is the typical way of helping students learn material is with problem sets. So they do problem sets, right?

They go to a lecture, then they get a set of problem sets and they learn how to do the problems. Then when it comes to the exam, they know how to do them and they do well on the exam if they did the problem sets. The temptation now to just go online and get the answer to the problem set has made it too easy.

So we need to think about how do we change our education system?

And I think the way it's going to change is more person to person contact, more of a model where students sit down in small groups with a teaching assistant and really go to the whiteboard and work the problem out on the whiteboard. I think that is going to play a bigger role going forward.

Lynn Thoman: You've thought a lot about leadership. Do you think the biggest constraint on human progress is going to be technology or leadership?

John Hennessy: The biggest limitation on human thriving, I think, will be leadership because good leadership ensures that all boats rise, that society as a whole thrives. Technology alone can't ensure that.

Lynn Thoman: So the obvious follow up is if AI starts outperforming humans on many cognitive tasks, how does that change what leadership means?

John Hennessy: In my experience, the hard thing about leadership is all the difficult problems are people problems. They're all people problems. And I don't see an AI system coming in and solving those difficult people problems.

I think people are going to still want to be talking to a person. It's the same thing when I think about the future of medicine. I do believe I want an AI system there reading my X-rays, reading all my diagnostic tests, putting that all into a model and trying to get some insight.

But if the doctor is going to tell me I have a serious illness, I want to talk to a physician, not to just the computer. So that's, I think, the way to think about it. And I think that's true in leadership, too.

People want to work with other people and they want to do things as a team that they couldn't do alone.

Lynn Thoman: Imagine a young person listening to this conversation who's excited and maybe a little anxious about AI. What would you tell them?

John Hennessy: I'd say learn how to use AI as a tool that will enable you in whatever field or career you go into to be more effective, because I think what's going to distinguish people in the future is whether or not they can use AI as an effective tool in the context of what they're doing.

Lynn Thoman: You're really saying don't use it as a substitute for homework or work?

John Hennessy: Right.

Lynn Thoman: You need that cognitive development, right? You need to do hard things.

John, what are your three takeaways? What would you like the audience to remember about AI?

John Hennessy: First of all, AI is a tool. It's nothing more. There's no magic sauce here.

It's a tool.

It doesn't know everything. It knows a lot of things, but it doesn't necessarily know everything.

And third, the most important uses of AI are ones that augment and enhance human capability, not simply replace humans.

Lynn Thoman: John, this has been wonderful. Thank you so much.

John Hennessy: Thank you, Lynn.

OUTRO: If you're enjoying the podcast, and I really hope you are, please review us on Apple Podcasts or Spotify or wherever you get your podcasts. It really helps get the word out.

If you're interested, you can also sign up for the 3 Takeaways newsletter at 3takeaways.com, where you can also listen to previous episodes. You can also follow us on YouTube, LinkedIn, X, Instagram and Facebook.

This transcript was auto-generated. Please forgive any errors.