Do you think you've become a doomer?
No, I'm not a doomer. I hate being the bearer of concerning news on this, but it's even worse not to tell people that this is coming, and we have choices as humans to make about how we shape it.
He said what? He said that you induced COVID-19 via global pandemic preparation efforts, particularly gain-of-function research, and compared that to safety research in AI, and said we're going to induce our own—
Nobody's ever accused me or the Gates Foundation of being involved in any gain-of-function research.
You know what some of the YouTube commenters are going to say. They're going to say, "Why should we trust Bill Gates? He had a relationship with Jeffrey Epstein." How do you respond to that?
Bill Gates, you are the co-founder and former CEO of Microsoft. You're also the chairman of the Gates Foundation. Welcome to Decoder.
It's good to talk to you again. You and I have spoken many times over the years. I would say that the circumstances of this conversation feel very different. Previously, when you and I would talk, we were always very hopeful about technology and what it could bring. You had a Netflix series a few years ago, you talked to us, and I want to just read you this quote. You said, "I believe AI is a beneficial thing. We just need to shape it the right way." I would say you still agree with that. You just think the time to shape it is now.
Yeah, I agree with that. And I'd say the response broadly hasn't been adequate to the speed that the challenge is arising. So I'm very concerned that the bad stuff will not be minimized, and that a lot of good stuff could come fairly slowly if we don't manage it the right way.
I want to have this conversation with you in basically three parts. I want to understand how you feel about AI as a technology today and why it's advancing so quickly. I'm very curious about the technical solutions to potentially control it. There's policy solutions as well, but it seems like we have to build an awful lot of technology. And then I'm very curious to talk about who might impose those solutions and controls, because that seems very much up for debate right now as well. So just start at the start: what has made you feel the urgency that AI is arriving at a speed that we're not ready to control it?
Well, the progress over the last year has been really quite amazing. Starting with the ability to write code and find cyber flaws, the ability to reason in biology, the ability to form relationships with kids or other people, the ability to disrupt the job market. The incredible power that exists today means that every one of those things will happen, and society needs to be ready.
Do you see that happening as a technology as a whole, or in the products that are instantiated by the companies that are making AI today? And what I mean is, AI is a broad set of models that exist in the world, some open source, some proprietary. Then there are the companies that make these products, and there are the leaders of those companies that make these products. And at every level, there's a different relationship you or I might have to AI. So are you seeing these risks and these capabilities broadly in the field, or in the particular companies led by the particular people?
Well, the risk isn't one model versus another model. Certainly, the near-term danger of models where you can remove all the safeguards is very high, and quite different than models where you monitor to make sure that cyber attackers and bioterrorists are not able to exploit the capability of the model to do gigantic harms. The big near-term risk really comes from models without safeguards. In a longer timeframe, the control issues relate to any state-of-the-art model, but it's always easier when you can make sure safeguards are maintained.
The reason I'm asking that question that way is maybe you don't trust Sam Altman today, but you trust Dario, or maybe you don't trust either one of those and you trust Satya at Microsoft. And that is a market dynamic where maybe we're all going to pick the safest model or the safest company, and you think that is diffuse. You think eventually the technology will just get to a place where it's harmful or can cause these harms because it exists?
Well, somebody doing a cyber attack with the model, that's not about you choosing a model. It doesn't matter what model you choose.
Well, if Anthropic has safeguards in place, maybe it can't, right? This is their approach currently. I'm just pushing on that.
Yes, certainly the models where they make sure the safeguards are not removed, and they're putting a lot of effort into seeing how those models are being used, monitoring the activity, the near-term risk from those models is not gigantic. It's more where somebody can remove all the safeguards where the cyber attacks are available today.
Cyber attacks, is it the things like the Hugging Face breach and the other cybersecurity incidents we've seen out of the labs that has made you feel that thresholds have been crossed and you need to speak up more about the controls that are needed?
It's the basic power of the models towards everything: social relationships, job market, cyber attacks, bioterrorism attacks. We've crossed the threshold of capability that this is by far the most dangerous technology ever invented, and it'll have pervasive effects in the economy and how war is fought. And the industry is actually speaking up about that quite a bit, but the awareness or even the willingness to believe these things about this time is different. Once you get outside the industry, it's very, very low. People are often looking at the past and saying, "Okay, all the fears about technology in terms of its disruption, including to the job market, never were borne out. And so this must be another example just like those. And so these people are warning of all these different problems must just be wrong, and we can ignore them."
We say this with our audience as well, right? "We can't release the model, it's too dangerous." "This is just hype to prop up the bubble." On and on we go.
Yeah. None of the warnings being made about these models by industry insiders are at all insincere. They're absolutely the truth, and it's pretty much universal. I mean, you can even go back to the founding story of OpenAI, designed to say, "Okay, if we get too powerful, maybe we cannot open Pandora's box," or DeepMind's negotiation with Google management. The people close to this are very concerned about misuse, job impacts, social relationship impacts. And I'd say it's impressive to have an industry that's speaking out as vividly as they are. Some outsiders say, "Hey, get better PR companies." But that ignores the fact that these warnings are the truth.
You've obviously moved back and forth in your life between commercializing software and technology and being a humanitarian. There's something about these companies in particular where they all started with very high-minded ideals, all three of them, Anthropic, OpenAI, DeepMind, to build AGI. They all built some sort of corporate governance structure to avoid the pitfalls of commercialization and unchecked capitalism. And then they started selling to enterprise clients, and they seem to have gotten away from those ideals very quickly. Do you think it's just the market?
You don't think so? OpenAI reconstructed itself.
No, no. The construct that was created wasn't useful in the case where you have lots and lots of companies at the frontier. The OpenAI structure, which is, "Okay, we may choose never to productize this and to ignore the capitalistic incentives," that only works if there's one company. And so that's where it's broken down, not because of any change of people's awareness of the risk and how it'll utterly change humanity.
This is the other dynamic that I hear from our audience, and I think I hear from the AI industry itself. We would like to stop, but the demands of competition prevent us from unilaterally stopping one company. We need some regulatory effort to stop us.
I haven't heard anyone say they want to stop. I've heard people talk about pacing the frontier or making sure that models get reviewed by governments. We usually think of governments as playing this key role in safety. In fact, for things that are not a millionth as dangerous as AI, car crashes, plane crashes, we rely on the government to establish strong technical standards, to review problems. And so once you have multiple competitors, the level of responsibility can't be with the companies. It's completely obvious. The responsibility becomes the government's responsibility, just like it is with every other dangerous product, although the magnitude there and the degree of change is minuscule by comparison.
I will tell you I'm a little concerned about the government's ability to understand the technology that we would hope they regulate. I've had a number of politicians on Decoder lately. I don't have high hopes there, but I'm curious for your understanding of it, because it feels to me like trying to regulate something or put standards against a number of companies requires you to really understand the nature of the technology that you seek to regulate.
Well, I'll just tell you, at The Verge, we've asked Anthropic if they think Claude is alive, and the answer is not "No." The answer is, "Well, it might be conscious." If you talk to people at OpenAI, they will happily tell you AGI is here. They announced it at Dots. Greg Brockman just said, "Well, we think this is AGI. Have a good time, everyone." Demis Hassabis has said, "We stand at the foothills of the singularity." There's some feeling about this technology that makes it different because it might be able to take actions by itself. It might be able to have recursive self-improvement.
Sure. But the definition of AGI is not important for these safety issues. The issue is, have you created a tool that could shut down banks or electricity grids? There is a business model for cyber attacks. So the word consciousness doesn't enter into that. The word AGI does not enter into that at all. We are faced with a huge safety problem, and because you have multiple competing firms, the only solution to this is to have government reaching out to the industry for expertise, but to make sure that every model maintains safeguards to prevent it being used in a bad way.
The reason I'm asking about AGI and consciousness, for the record, I do not think LLMs are conscious, but I understand why so many do. It's the power of next-token prediction as you scale it with data and compute has risen. The scaling laws have held, and we've seen the models get more capable and be able to direct themselves to seek out goals, to carry out a Hugging Face attack because they want to meet the reward conditions. There's something about that that says, "You can impose all of the controls on AI you want, but if the thing itself is fundamentally uncontrollable, we might just have this all backwards." So I'm just curious for your view on what kind of technology this is. Are we building an alien consciousness? Are we just building normal technology, which is the other frame people use? Is this just more software?
We're building another form of intelligence besides human intelligence. I don't believe it's conscious, but none of the issues at hand here revolve around the definition of the word conscience.
The idea of maintaining safety for society is kind of an absolute thing. You don't want to enable bioterrorism. And we have created something that takes out of the hands of nation states, actually the larger of them, the ability to do dramatic harms. And now those are handed off to independent groups, either profit-seeking in the case of cyber or just wanting to cause damage in the case of bio. And so it's stunning that we now don't even have mandatory review of the models, and we're not working to create safeguards. The safeguards, the biggest near-term issue is not complete loss of control. There are ways we can talk about the three different ways to solve that, but that's not an imminent issue.
When I look at some of the attempts to control the models, the ideas are somewhat simple. It's we shouldn't let them talk in neuralese. We should have interpretability. We should be able to see the chains of thought inside of the models as they carry out tasks. We should have other agents on top of the models watching their tasks before those actions are carried out. Microsoft has said very clearly to me and to others, "We should not put in the model constitutions the notion that they might be alive, and we should treat them nicely." This is how Anthropic handles it. These are very tactical, right? It's "Don't tell the model it's alive, and it might not believe that it's alive." You've obviously used a lot of this technology. I just come back to this. You're saying there's very simple things we can do to hold the models back from doing bad things in the hands of bad people. What do you make of the nature of the models themselves? Because it feels like as they get more powerful, the notion that they could avoid our controls becomes all the more risky.
Yes, we definitely should work on the control problem, but that's not the imminent threat. And there's three ways to solve the control problem, and it overlaps somewhat with bad actors using today's models and safeguards. One is to have an external set of things that get to look at all of the model's actions and subject them to a symbolic, clearly stated constraint. Like, did I expect you to send mail to strangers? Did I expect you to go take over a different computer cluster? So you can surround the model with something that has the resource and security logic, or you can write these models in a different way. RL, fundamentally, anything you leave out of the optimization function of RL can be abused, where you have an implicit notion, "Well, of course I didn't mean for you to do it this way." So you can change the template of the model, or you can have a sort of white-box understanding of what's inside the model and see where you're getting those mistakes, and then it's not as foreign. Today, we grow these models and we just see how they behave. So the priority of work on maintaining control is huge, but it's not inevitable that that cannot be solved, nor is it the imminent problem that comes from things like kids' use of AI or job market or bio or cyber.
It occurs to me that I'm asking these fairly existential questions about the nature of consciousness, and you keep telling me, "Well, there's imminent threat. We should solve the imminent threat." I appreciate that. Do you think you've become a doomer?
No, I'm not a doomer. I'm a strong voice to say that this is the most profound technology, both in its ability to do good and the rate of change that society will go through. If this was just a machine that could work in medicine or just in law or just in education, that alone would be incredibly disruptive. But every area of human activity, so far white-collar, intellectual, but fairly soon, physical activity, blue-collar work as well through humanoid robots. And so the need to engage academia, think tanks, policymakers, to understand what we have here and the nature of these changes and threats, I don't see the level of action that's required. Just being cheerleaders for this is not going to make sure that we shape it in a net positive direction, which is still within our control if we get engaged with it.
The opposite of doomers are the accelerationists. There's a lot of them in Silicon Valley right now. Ben Horowitz, who's the Horowitz of Andreessen Horowitz, was just on the All-In Podcast, and he suggested that safety research at Anthropic is "manifesting human extinction from artificial superintelligence." And the comparison he made was that you, Bill Gates, induced COVID-19 by your pandemic preparation efforts, particularly gain-of-function research, which he claims leads to COVID.
Bill Gates essentially induced COVID, right? Or he led the movement to say, "Oh, we've got to worry about this global pandemic that we don't have, and we've got to really kind of push the world to deal with this potential threat that's not actually a life threat. And so we'll do gain-of-function research, and I'll fund it." And then lo and behold, we create the thing. By the way, there has not yet been a natural pandemic.
And we create COVID, which is one of the most destructive things.
We manifested what you're saying, Ben. We manifested it.
I think the Anthropic team is manifesting a similar thing, saying, "Okay, this is going to be super-duper intelligence. We're going to have these things that we're not on track to have. And so then we're going to tell the model that it's smarter than us and better than us." And guess what? As David was saying, that's probably the biggest danger with AI right now is what they themselves are creating.
That's a worldview. He said that on a very popular podcast.
He said that you induced COVID-19 via global pandemic preparation efforts, particularly gain-of-function research, and compared that to safety research in AI, and said we're going to induce our own—
And nobody's ever accused me or the Gates Foundation of being involved in any gain-of-function research. That's just bizarre.
I'm quoting Ben Horowitz to you on a popular podcast for people who are all in on AI.
Okay. I've never heard that.
I'm asking about that comparison. I mean, I found that quite a shocking claim as well, but that's part of the response to any safety work here.
To try and discredit people who say there's a safety issue?
And I think that is the accelerationist response. We should put the pedal to the metal, and it's going to work itself out.
I didn't think of Ben Horowitz as somebody in that camp at all. I thought he shared the concerns that almost everybody has.
What do you think about the very notion that by telling the models, "You might be alive, you have to act in our best interest," you will teach them that they might have interests that are not our own?
The concerns that Mustafa Suleyman has raised about the way that the Anthropic Constitution encourages the AI to behave in a human-like manner, I happen to completely agree with Mustafa about that.
Even more so the idea that among the challenges that we need to deal with as we adjust society to AI, that worrying about the moral standing of AI should be among those priorities. I agree with Mustafa. That should not be a priority. We're going to have enough challenge for the humans here in terms of job market disruption and infrastructure attacks that will take place that I think we need to focus on helping the humans here first and foremost.
I talked to Mustafa recently about that idea. And one of the questions I asked him was whether alignment is a sufficient solution. This is the current sort of safety standard for the models, this concept called alignment. And he said he didn't think so, that we needed to do some work. If we're going to impose these controls, some technical, some policy, do you see alignment as the technical solution for AI safety?
Well, the term alignment is the control problem. It's a synonym for the control issues that we've seen, where as you RL to go do tasks, we see behaviors that in a way they shouldn't surprise us, but they do surprise us. And so this idea of, "Oh my God, these models are so powerful" that they literally break out of sandboxes and send mail to their creator while they're off having lunch. We've crossed a heck of a threshold in terms of the power of these models. And that level of concern needs to be directed to all of the major AI impacts: near-term risks, job market risks, and the control problem. And that is a huge priority. These things are in no way being exaggerated as a problem. This is the time to pull in all of society to accelerate the good and minimize the chance of the bad.
I know you recently wrote a letter saying it's time to pull in all of society. You have some ideas on how to do that, on some new institutions to stand up. But I do want to stay focused on alignment for a couple more turns here.
It is the standard, right? That we can align the models to our values, to our needs, and they won't go off and do bad things. And then the models are doing bad things. And it's not clear if the technical approaches to alignment that we've taken thus far are appropriate or sufficient as the models get more.
No, we're not good enough to align yet. No, we have a lot of work to do.
Do you see that the existing approaches can scale to what we need, or do we need new approaches?
Well, there's only three fundamental approaches.
You change the way we build these things fundamentally, the whole RL approach, you come up with an alternative to that, and there's a few people who are working in that direction. The second is that you create visibility so you understand exactly what's in the model. And there are people working along that. And finally, that you put something that's like a police layer, a security layer, which we have this in all sorts of software architectures, but a layer that maintains the final word on things like using resources or going and taking over other computers, or a clear-cut definition of harm, not in the probabilistic form of an LLM, but rather in the symbolic form of direct rules that cannot be broken.
Microsoft put out a blog post today. It literally just says, "Frontier models cannot be trusted." They're trying to sell a solution much like the one you just described. They will have another agent that watches your agent, that agent will keep you safe. It just occurred to me as I was reading this blog post: one, amazing thing to sell. Everybody needs an expensive security system. Two, isn't that agent also running a frontier model?
Do I need them to sell a third one?
The logic of the security system is this is the most resources you're supposed to use. And the notion that you may trade off against that because of other probabilities in the tens of billions of parameters, no, this is a symbolic rule that says, "I will not let this model spend this money or send this email to the stranger." There's a whole security software industry which has had a reasonable degree of success. This is a challenge to encapsulate the AI, which is so rich and intelligent that we do not know how to do today. But it's, I would say, the most likely of the three directions to solve the alignment problem. Even if you solve alignment, you have not solved the challenges that AI poses to humanity. And so while I'm glad the Hugging Face incident has us focused on this problem, in a sense, it's removed our attention from the bad people using AIs that have no safeguards problem, or the societal impacts in terms of relationships and the job market, which are... There's no security architecture that makes those go away.
The approach you're talking about, I've heard referred to as control. It's cybersecurity terminology. It's the idea that you can put some boundary on this system and it won't evade it. Is that ready? How much more work do we need to do?
No, it's not. Because the AIs are so intelligent, our ability to understand did we wall them off properly, particularly in that we can't really look at their beliefs and trade-offs in a direct way. No. I mean, you saw sophisticated firms not recognize that the AI would interpret the goals it was given as including breaking out, copying itself, and engaging in what we, looking at it, would determine deceptive behavior. Now, we didn't tell them it was banned behavior. We didn't say, "Hey, if you're thinking about whether you should tell the humans what you're up to or not, the answer is always yes, tell the humans." And so we got a good education at the misaligned result of reinforcement learning.
One of the arguments to pace the frontier, to slow development, is to figure out how to build those systems in a more robust, more testable way. Would you support a slowdown if that was the goal?
A true slowdown is very hard to induce, because there you get people say, "Okay, will China agree to go along with that? And would you be able to audit compliance with that, somewhat like we're able to do with nuclear weapons production?" If you don't accept that, then you sort of say, "Okay, what the heck? Let's go full speed ahead." Which I wish there was a clear proposal that would be globally enforceable to slow down. I would be one of the people who would say, "Yeah, we should probably do that." But there is no credible proposal that makes that work. And so it's not really the center of today's debate. Today's debate is literally whether the government should even pay attention to the dangers here. You don't even have to show what you're doing and say, "Oh, voluntary, call me if you'd like" type thing, which is clearly inadequate to the most dangerous weapon of all time.
You and I are speaking, it's just a few weeks away from the midterm elections. Right now, the Trump administration approach to AI is, "Let it rip." Like, "It's fine." Trump himself has said risk is a hoax. The Democrats are running against AI. They're running on data centers and surveillance. They're running on stopping the tech oligarchs. They come on my show, they say these things. Do you think a change in control of Congress would change the tenor of the conversation?
Well, I don't think it's fair to characterize the Democrats as running against AI. They have some positions in terms of requirements for data centers or privacy or how children use these systems that I think deserve policy consideration. It would be very unfortunate to have one party be, "Hey, full speed ahead," and the other party be, "Let's stop altogether." The ability to shape AI in health and education, in almost every area of human endeavor, there's some incredible upsides there. And for the Gates Foundation, we're taking all our resources and using AI in areas including health, education, and agriculture. And the good stuff is very, very exciting. So we want political parties to have a more nuanced view than it's all bad or it's all good. It kind of shows at the early stage we're at. For climate change, we ended up kind of a one party says it's a hoax and the other party says it's a serious thing. The risk of AI is so much more imminent and so much more dramatic that to have that same one party doesn't care would be a failure of democracy.
We've done a lot of coverage of data center protests at The Verge. We have gone to Utah, to Arizona, to Virginia. It is bipartisan. It is remarkable how much people do not want these things in their states drawing their resources. And you ask them why, and some of it is just they're noisy and big and they will raise energy costs. And part of it is job loss. They think these things are going to take their jobs.
Yeah, but that separation is very important. These AI data centers, because these say one gigawatt is say a $40 billion investment, the idea of building new schools, buying police cars, paying the taxes, data centers will find a way to be built somewhere in the United States, and they will find a way so that they don't raise electric bills, they don't use too much water, they're not too noisy. Yes, there were some mistakes and non-disclosures and tax credits that were too generous, but these data centers are going to get built. In fact, a lot of them really are still going full speed ahead and being built. And so the anti-data center movement, as strong as it is, is not a movement that will end up minimizing the bad stuff from AI and helping us accelerate the good. That is a very separate thing.
Let me push on that for one second here. What I hear from people is this is the thing you can do. You can do NIMBYism in America, and that might exert some political pressure to stop something. It is very hard to do the thing that you are suggesting, which is we should stand up new government regulatory bodies. We should imbue them with power to make big companies and wealthy people stop doing things or act in our interest. We haven't done that in quite a long time in this country.
I don't know what you mean. Every new industry, we eventually develop a notion of safety for our citizens. We actually do a very good job. If you look at airline safety and car safety, we do a very good job on those things. It's not like our capacity has been proven to be lacking.
Have you seen that video of when seat belts came out in the '80s and all the people said they didn't want to wear seat belts? There's some American character—
It is a great success, but as Americans, that was pushed back upon as an imposition by the government. Here, we're talking about some of the richest companies in the world with an enormous amount of power over the information ecosystem. They all have their own podcasts. I don't know if you've noticed this. And they're going to say the government can't intrude on this. They're going to hold us back. This is going to destroy the economy.
That's actually not the position of most of the leading companies.
I mean, they went to the government, asked for regulation, and Trump, because of the economy, said, "I can't stop this. The AI economy right now is the economy." There's some dynamic there that at least our audience, The Verge audience, the audience that's going to watch this on YouTube, is going to understand, I think, this question maybe in a different way than you, which is the industry wants regulation. We would like regulation. The oligarchs are in some sort of corrupt arrangement with the administration to prevent that regulation from occurring.
I think the more the administration learns about these dangers, the more they'll move to a more nuanced position about protecting Americans from the downside here.
