Bill Gates is a fascinating person in the AI debate right now. He is somebody with experience in a number of the different perspectives that most people can only hold one of.
Hello, I'm Bill Gates, chairman of Microsoft. In this video, you're going to see the future.
Here's a revolutionary technologist who built some of the foundations of the future that we're now living in.
What can the computer do that a book can't? I mean, that's the foremost interactive process, isn't it?
No, it's not. Not at all.
An innovative word processing program called Microsoft Word.
This is absolutely revolutionary. Of course, then of course as CEO of Microsoft, Microsoft, the world's largest software company. Some say it is the most successful company, the most dazzling new industry of the century.
He has felt the momentum of corporate competition. And Microsoft of course is still in some of the race dynamics present in AI.
We now have an incredible roster of seven new world-class models to keep everybody working at the absolute frontier. And then as chair of the Gates Foundation, he has been working with governments around the world on regulatory issues, on poverty alleviation, on equity for many, many years now.
Very, very few people combine technological experience, corporate experience and governmental experience in quite the way he does.
Certainly computers will be in any meaningful sense as smart as people at some point.
So his recent essay on AI where he says now that he is staking his reputation on trying to get people to see how bad what is coming might be, trying to get them to see that we are not ready for what is about to happen, was something, it was a real departure from what I've read from Gates previously to this. And then when I sat and talked to him, how emphatic he is, how afraid he even seems to me to be that we are not ready for what we are building and that so many of the people in positions of authority are denying what is about to happen, was really striking. It's a really quite a call to arms.
Bill Gates, welcome to the show.
So, I wanted to begin with a clip we found of you on the David Letterman show from 1996. I'm going to hand it over to you to play.
Is there something now beyond what we understand about computers that like 20 years ago we didn't fully understand computers? Is there now another level of something? Maybe we haven't even thought of it. Maybe it's not even possible. Maybe, you know, a whole different mechanism, a whole different software and hardware, or is this going to be it now through the end of time?
Mostly what we're working on now is the computer being a tool, a tool to help us learn or find other people with the same interests. Eventually we may figure out how to make the computer think, but that turns out to be a very tough problem. In fact, there's been almost no progress made on it. So, nobody knows when that'll happen. Some people think it'll never happen.
Yeah, we don't want them to think, do we?
Not really, I would think.
So, that was 30 years ago. Narrate for me how we went from that being a scary thought that might never happen to arguably the reality we're sitting here discussing today.
Well, the notion that computation could provide thinking at a human level. You have Alan Turing talking about that before I'm born and even proposing a test of if you could be fooled in a conversation, that was called passing the Turing test. And so the whole time I'm, you know, learning software, this idea of can we make software see or listen or read or write, you know, that's the holy grail. And when I do drop out of Harvard, I said to my co-founder Paul Allen, gosh, if there's a breakthrough in artificial intelligence while we're off, you know, selling BASIC interpreters and word processors, you know, we'll feel bad. I might wish I would have stayed in academia. So that's 1975. Progress on AI is mostly going down dead ends like Prolog, expert systems, and there's a small group including Geoffrey Hinton and a few others that are working on neural nets. And eventually there's enough power actually coming from the graphics processor that that idea, these highly statistical approaches, start to show promise. And I was going down to see OpenAI on a regular basis to see the work they were doing. And I challenged them, hey, if you can read a biology textbook and pass the advanced placement exam, getting a perfect grade, which is a five, you know, then you will have proven that you are reading, that is encoding knowledge in an accessible form. And so it's six months before the public release that Sam and Greg and Ilya come up and at my house demonstrate to me getting a five on the AP exam. Even on questions that I had made up that it couldn't possibly have seen, very complex biology problems, it was nearly perfect. And so that was, you know, shock number one. And then late last year when the Claude, Codex models have gotten super good and I can see that they are as good as I am, you know, which is significantly my most developed talent because I was obsessed from age, you know, 13 to 24 that, you know, could I write code as good or better than anyone, that's another moment where I go, this is incredible, that the capability of doing long-running complex tasks has now gotten to the point that they are superhuman at writing code and in finding flaws in code. So what they're not superhuman at yet is deciding what to do, that kind of higher level strategizing.
Do you think that's far from being a capability for them?
Well, definitely. If you're the, our foundation does these strategy reviews. We spend two weeks in October to set, you know, how we're going to spend our 10 billion next year for 2027. And a year ago, somebody said, 'Well, we should ask the AI what it thinks.' And that was actually a pretty good joke back then because it wasn't coherent enough to see these things. This year amongst the inputs we'll have to that discussion is taking strategy notes and actually engaging in a dialogue with ChatGPT, Claude, Copilot, and even having them talk with each other. And in a few of the reviews we'll actually have them sit in. In a few cases we'll tell it, hey, only speak if we ask you, and then in a few other cases we'll say, hey, if you hear something you think is wrong or you hear us thinking, what are these statistics, please engage. So, you know, we've gone from it being a joke to it will be a peer, not making any final decisions, but it will be a peer in deep complex strategic discussions making a significant contribution.
So, in that Letterman interview, you said it would be a scary thought. Why then would you have said it would be a scary thought to have computers that think?
Well, no one who's ever been fascinated or wanted to develop AI doesn't realize that it's incredibly scary that it will be better than biological minds. You know, it's amazing how general purpose they are in that the optimization was staying alive, breeding, socializing with each other for survival and fertility. And yet, you know, we can write symphonies and play chess and, you know, even write some pretty cool software. And the idea when you move the template away from biology to silicon, you don't have these boundaries between individuals. You don't have a limited memory. You know, the size of the brain is limited by the birth canal. It's, you know, why humans at first are very limited, we're very unusual in how helpless we are at birth because we've so optimized for having a large brain. But the silicon intelligence doesn't have these limitations. The idea of read every medical journal and see if there's anything that we didn't spot, the AIs do that today. That's why particularly for less common diseases, they are so superhuman at seeing a set of symptoms and being able to diagnose them. They can just keep more in their mind and see what things relate to each other. That is, you know, no human will ever be able to do that. So that if you don't retain control over it, you've evolved a species that will be to us, you know, as we are to say, you know, dogs or cats, just in a very different realm. And so every AI company, whether it's DeepMind or OpenAI, you know, they all say, 'Okay, whatever goes on here, it can't just be driven by profit maximization. We have to have a charter that if we get to dangerous thresholds, we can exercise judgment that would be against profit maximization.' Sadly, those mechanisms only work if there's only one company. And, you know, so say OpenAI invented, you know, post-AGI and then they said, 'No, no, we're going to bury this.' If no one else ever did it, then fine, that Pandora's box stayed closed. But of course, many companies work on this. And, you know, even OpenAI, Anthropic, because they think that some of these safety issues aren't getting enough attention. So no one involved with this takes lightly the idea of, okay, what world does super powerful AI create?
I want to hold on that race dynamic for a minute. One of the reasons I was excited to talk to you about this is you've both been on the technology side, but you've run a major company in competition with other companies. You've worked with a lot of governments. I was out, I spoke last week with Jensen Huang of Nvidia and he said that his perspective is that safety is a real concern but the race dynamic is fake. If the product isn't safe, don't release it. If I believe that I'm about to launch a product that is unsafe, it is completely in my ability, my power, and my responsibility, and I'm incentivized to do so to not launch the product. This is, we don't need new laws. This is the role of individual CEOs to not release a product that is not ready to release. How do you see that question?
Well, there's never been a product that's less understood in terms of what its capabilities are than AI. And AI has crossed the threshold that its ability to empower a bioterrorist to kill hundreds of millions, that exists today. The ability to do a cyber attack that scrambles all the bank accounts, shuts down the electric grid, that exists today. And we know that's the case. And the reason that exists is because somebody with ill intent can take the AI and cause it to do those things. And it's not the AI, you know, someday in the future, the control problem where the AI on its own through an unintentional interpretation of what it's optimizing could go off and do bad things. But we crossed the cyber threshold and we crossed the bio threshold early this year. And my decision to take my voice and not just be, hey, let's eradicate polio, let's be generous with foreign aid, you know, the things that all of my money is going to that I am going to use my boost, it's something that's more important which is we are not awake to where we are with AI and choices that humanity, not a country, but all of humanity has to make, that do we make the effort to shape this in a net positive direction, actually overrides my total commitment to the foundation health work.
What specifically was the threshold or what did you see that made you think we're in a new reality here?
This notion that it can find bugs in code, including security bugs, the next releases leading up to Mythos are increasingly good and they're finding bugs in code that humans have looked over for over 20 years and said, 'Boy, we see there's no problem here.' And it in a few minutes says, 'No, no, I can inject this over here and this over here.' At a level of complexity where when you see it, you go, 'Wait, oh yeah, you're right.' And so the cyber hacking capability was stunning. And then there was this notion called Glasswing that you would give it to a few people so they could try and fix bugs before it got used. But there's way too much code and so we're just in a period of extreme vulnerability to cyber attack, the bioattack. You know, the Gates Foundation, we fund lots of medical research so the sophistication of coming up with new molecules, that's really a good thing but you know, it's ultimate dual use because if you want, you know, something that's say worse than smallpox that it takes even longer to show symptoms before you're infectious. So, you're infecting a lot of people before it damages your health. It used to be only nation states had enough resources and capability to do these things. Now, that power has been passed into the hands of a small group just using the latest AI tools.
So why isn't it enough to just say listen there is product liability now you release a product it helps some terrorist group create a bioweapon that's going to be very bad for your company you're not going to do that right and they have categorizers and other things meant to stop people from using bioweapons we've begun to see the beginnings of control issues with things like the Hugging Face hack where you know at least experimental AIs are breaking out of sandboxes and coordinating to do things that are way outside the scope of what we would want them to do but again those are non-release systems you know Anthropic you know withheld Mythos trying to create more cyber security. So why is anything needed beyond and is anything needed beyond the simply natural incentives under capitalism and normal corporate reputational management?
Well, I almost can't believe you're asking that. This is the most dangerous thing that humans have ever gone near. In other areas, do we just say, 'Hey, release your drugs. There's no FDA. There's no airline safety board. There's no requirements that cars use seat belts. Do we just use the liability laws to try and keep humans safe?' You know, oh, you're shipping opioids. Somebody should just sue you. I mean, we've created a society that tries to keep people safe. Not by saying, 'Oh, we can bankrupt the person who does that.' The harms here. And you say there's filtering. There's not filtering. You can take an open-source model that can create bioweapons and disable any monitoring of any kind and this exists today. So no there is no filtering of any kind and so, you know, say, you know, you kill a 100 million, you want to use a lawsuit, I almost can't keep a straight face.
Well this is not my view but it is President Trump's view it is David Sacks's view to some degree. It's Jensen Huang's view and so that's why I'm putting you in conversation with it because it is the governing view of the United States of America at this moment.
No, it's fair to say that outside of the industry, the awareness of the dangers of AI is extremely low. And you can say that of academia, you can say that of think tanks, you can say that of policy makers, politicians. And part of the reason that, you know, I'm speaking so loudly, as loud as I can, is that you can't rely on the industry to self-regulate here. I mean, it's just insane. The only question in my mind is do we wait until a cyber attack causes massive damage and a bioattack causes massive damage and then the monitoring safeguards are required in these models to minimize the chance of that happening many, many more times or can we be wise enough to put these things in and require these things to be put in before millions of deaths.
The two things have been worrying me most as I'm tracking what I'm hearing from people in the labs is one the view that these systems are becoming less monitorable as it becomes smarter. OpenAI said this about Astra 6 and people inside OpenAI have been raising the alarm that the systems seem to know when they're being tested. There's more situational awareness and so the testing environment is no longer as indicative necessarily of what they will do. So that's one. And the second is that all the labs are moving as quickly as they can towards recursive self-improvement where you have the AI coder moving not just from speeding up the engineer but really taking over the coding right and so you can create a much faster iteration loop with much less human input than you have now. So when I think about those two things and when I think about what I'm hearing from people, my confidence that we have the monitoring, the auditing, the testing capabilities we are going to need is low. I'm curious where yours is.
Well, so you're taking, you're not focusing on the whole problem there.
The fact that at some point the AI system itself may do something that's against our interests, that is absolutely a risk and, you know, the symptoms we've seen through aspects of Hugging Face and other problems point out that the way we do the reinforcement learning today, the fact that we don't have a supervisory layer that is, you know, some absolutes like you're not supposed to take over other computers. You're not supposed to break out into the internet. The control problem is a serious problem, but the imminent risk is not RSI. The imminent risk is these are the most powerful tools ever. And unlike every other dangerous technology, they were not funded by government R&D. And the government is not a significant purchaser of these products. It's not like rockets or nuclear weapons. And so the idea that the government including, you know, the US government that would have traditionally been the most technically sophisticated other than a little bit of attack capability in NSA really doesn't see how dangerous these things are. That is a unique circumstance.
So what does that imply? If you're more worried in the near-term about what human beings do with AI than loss of AI control, then what does that imply of what your first steps need to be?
There is no supervisory layer today. That supervisory layer is needed urgently to prevent bad people from using today's AIs to shut down economies or kill millions of people. These models, the ability to separate out making molecules for good and making them for evil, that is so hard to distinguish. You know, Anthropic released Mythos and Fable where they had turned the filters up so high that you can literally ask questions about cancer and, you know, next thing you know, you're down at Opus and then Sonnet and then, you know, Haiku is the only AI willing to help you. And so it's overtuned and anybody who wants to do serious work either has to get a special permission copy or go use something where there are absolutely no safeguards at all, which include some of the open-source models. And there are humans who for bad reasons will be able to use these models and so unless we put in safeguards and monitoring and require that in all models then we're just going to have some gigantic events of that type and then finally we'll respond. Sadly, you will have let AI capabilities go off in dark places even beyond today's capabilities which that is a gigantic mistake.
So, you just wrote this essay on AI risk and it does represent for you a big jump in how alarmed you sound, right? I was reading some of your past pieces. The one in July of 2023 was titled 'The risks of AI are real but manageable.' And here you're I think at a different level of threat. So I take from what you're telling me that what happened here is just watching the advance in bio capabilities and cyber capabilities and coding capabilities etc. Is that fair?
No, the key thing is we always said when we cross these thresholds we will engage all of society because we will have created the most dangerous thing ever. This makes nuclear weapons look like nothing. And we said we would engage and we crossed those thresholds and there was complete silence. There were discussions about who in the industry says X or who in the industry says Y but, you know, what percentage of academia or think tanks or, you know, what, so it's the complete lack of response. You know, I would love a world where the response on these things is very strong and then whatever, you know, time or voice I have will go back to let's eradicate polio, let's stop children dying under five. But it's the combination of having crossed every dangerous threshold without a doubt and a complete lack of engagement outside of the industry.
Well, it doesn't feel to me that we're so disengaged. This, I mean, as somebody who covers politics, every politician I know is talking about this. There's bills being proposed. You know, Donald Trump, say what you will,
Not to have to do with these risks.
Well, say what you will about Trump and them. They're engaged. They just believe we should move forward. I mean, they're a little bit all over the place. One, at one point they're, you know, withdrawing access to things like Mythos and Fable. At another point, they're saying we have to win the AI race with China. But what I mostly see happening is a lot of engagement.
Well, when I talk to the people, and I bring similar concerns to what you have, they say the most important thing is that we win the race with China. Secondarily, many of the people with a lot of influence say the most important thing is we don't restrict access to open-source models. You know, you're going to need wide dispersion of them. And so the cost of regulation, the cost of the kinds of closing down of access, it's simply too high. The technology is still under-formed that we're sort of just trapped in this dynamic and the best thing to do is to just move through it. It's not so much that they feel disengaged to me as they feel they've come to a different conclusion.
So they don't mind bioterrorism?
I think you've talked to Donald Trump more in the past couple years than I have.
But this, you know, the last time I talked to him in December, this was not the big issue. We had not crossed the thresholds and we hadn't seen this. You know, the necessary step is people can talk about whether slowing down is good or not but putting in the safeguards and the monitoring will not meaningfully slow things down and the only effect that has on open source is that you can still be free, you can still be customized but you have to stay on a platform where the sovereign can make sure you have not removed the monitoring and other safeguards. And so, you know, what is the downside? The downside is staying on a monitored platform. Now, the get out of jail free card that people play here is that means the Chinese will win. I don't know what it means win. You know, nobody's, the US can't win over China and China can't win over the US. We both have opened Pandora's box. It's there. It's there for, you know, people with mal intent to go and use it. So you'd think we'd move up from that nationalistic view to kind of a humanity level view about how we engage in these protections. And, you know, the notion that China wouldn't want to engage in that on behalf of humanity, I disagree. And, you know, it's a proposition to be tested. If you can take models that make dangerous molecules and move them into a dark corner where you get rid of all the monitoring, then we're just giving up to a bio attack. Likewise for cyber. If you can remove that ability to find and exploit security problems which are rife then in the next few years you will have major, major cyber and bio events and that's avoidable through safeguards and monitoring.
There's another layer of risk that your essay talks quite a bit about that AI really will take jobs from people. Now I know a lot of people debating this who say no the jobs are too messy. There was a prediction we'd have no more radiologists. We still have radiologists. AI can code but we still have coders and strong demand for coders. So tell me about why you think AI really will take jobs away from human beings at a significant scale.
Yeah. So there's no doubt to date AI has created more jobs than it's destroyed. The demand for the skill sets even just to build the data centers is very, very high. We have a reasonably low unemployment rate. The superiority of these systems is subject to a threshold where you have to believe that it's incredibly reliable. You know, if you're going to have your telesales or telesupport capability be AI-driven, you want to make sure that its accuracy is better than humans. And so the only profession we've truly crossed over that threshold is coding. And even there, I know a lot of people who are, you know, kind of stuck in the past and, you know, don't want to use the AI for coding. But managers of coders get that most applications can be developed very inexpensively. We will in the next few years cross over those thresholds for accounting, legal work, telesales, telesupport where 24 hours a day in every language with infinite trivia capability and no urgency, you know. So you look at the AI nurse and there's several companies. They have perfect memory. They don't hurry.
Be like a nurse you would call on the phone or talk through a chat interface.
Far preference, you know, you go to San Francisco and ask people, would you rather ride in a Waymo or rather ride with the human driver? Go ask people in the UK who use Limbic for mental health support. You know, so the notion that just the market demand preference for driving or the nurse, you know, or even the person on the phone will favor humans, that's a quality threshold which will be passed through. You know, if say insurance companies are still doing humans through claims, medical claims, which is a very AI-capable task, then a competitor who has very few human employees will come in and change the pricing model for that industry.
So the two counterarguments I've heard people make on this one is that you have a Jevons paradox effect where the cheaper more widespread availability of this kind of intelligence leads to a massive increase in the demand for this kind of thing. So yes, you have many more AI chatbot nurses and that leads to more people being sent to the hospital, being sent to the doctor where real nurses take care of them. Or you have many more AI coders. So maybe my small podcast team, which wouldn't have had a software engineer before, now has one because it's like they run a team of coders and we can build products we never thought of before. This is the most common answer I hear to this that yes, AI will destroy jobs. Yes, it is making a human-provided resource much cheaper. But because it's going to expand the demand so much, it will sort of work itself out, create jobs in other areas, create new demand, and this is how past technologies have gone and so we shouldn't worry too much about this. How do you see that?
Well, Jevons is just a referral to the fact that parts of the economy are subject to demand elasticity. And yes, in the case of software, if you're say three times as fast and there's still some role that only humans can perform, then as you lower the cost, you induce demand. And so it's fair to say that the equilibrium today for software is not a loss of employment. When you invent radial tires that last four times as long, for some weird reason, people didn't drive four times as much. And factories that make tires employ a quarter as many people. When you replace people in Amazon warehouses with robots, people don't buy more because of that. So, you know, anybody who's numeric can say to themselves, what portion of the economy is subject to demand elasticity and what are those tasks that will still be human necessary? As soon as you complete the entire task, it doesn't matter that there's demand elasticity. That goes into the token budget. It doesn't go into the human salary budget. The cost for some of these things is so much less than the cost of the human labor. And so as you cross reliability thresholds both with white collar and human robots, you destroy jobs and you leave no high ground. That's, you know, innovation in the past. You have a tractor, fine. You know, let's build Disneyland and employ a lot of people there. You don't have that in the broad economy. You know, the people who say there'll be net additional jobs, I don't understand what they're thinking. They must not understand the piece of improvement we're on that, you know, the reliability.
You wrote a column I think was in May that I looked at like what, what is this unique humanist thing that you think, you know, in every category where AI has come along the preference for the AI is very, very strong.
Well, so that is based on so Jevons paradox and then the other which is an argument from Alex Imus who I believe is now at one of the labs as an economist is that you will have this sort of explosion in the relational sector that one thing that happens when people get wealthier and you'd probably know about this is that they all of a sudden get a lot more human help, right? They have personal trainers and chefs and there's a lot more.
Royalty used to have a lot of human help. You had upstairs, downstairs, you know, all those maids and people and you used to have a human who helped you get dressed. Believe me, the labor intensity of wealth is down super direct from what it used to be.
Jevons paradox. Name a blue-collar profession that's subject to Jevons paradox or do you not care about blue collar?
I do care about blue collar. So the question is anything in the blue collar realm subject to that.
So I'm not, I don't have the strongest view on this but here's I think the argument that I would make or that I've heard made is that if you look at something like manufacturing, we've not had a Jevons paradox in manufacturing employment in America. Right? We exploded how much we actually create, but a lot of the jobs went offshore. Some of the jobs went to automation, but we don't have less total employment in America than we did in 1960. And the reason is people moved into service sector jobs. That the composition of jobs across the economy changed.
Yeah. Human cognition became the scarce element. It's always about scarce elements. Once you replace human cognition as a scarce element, in fact you blow it away in terms of working 24 hours a day reading more knowing every language there is no scarcity that you're moving up.
So you feel there is no scarcity that will be left for human.
The Alex would say it's something like relational sector jobs but you don't believe.
Of the current jobs you mean like my relationship with a cab driver or my relationship with that nurse where I'd rather have Hippocratic AI call me. I think the question here is actually do you end up preferring Hippocratic AI or do you actually want your therapist to be a human being because after a little while there's something thin about telling your problems to a computer.
Well, the market, you can gather market data if that's at all interesting. You know, I suppose telling the 55-year-old truck driver you're going to go do some relational thing. You have a program for that?
I don't have a program for it. And I think the question you actually run into very quickly is speed of transition. Even if you believed some of this and I said this in the column too that I don't see how you will handle the speed of transition that our actual experience with fast transition of jobs is that people lost their jobs lose out. Now you might over the entire economy it doesn't look that different but say the China shock did lead to a lot of ruined communities. So your answer to this and it leads to a very concrete idea for you which is that we should tax the use of AI. Why and how?
