← The AI Policy Podcast
The AI Policy Podcast
A Crash Course on AI Standards with Google DeepMind's Owen Larter
00:38:44
Owen Larter, Head of Frontier Policy and Public Affairs at Google DeepMind, joins to explore the often-overlooked world of AI standards and the role they play in shaping how AI is developed and governed.
Loading summary
Transcript63 lines
- [00:00]
A
Foreign. Welcome back to the AI Policy Podcast. I'm Gregory Allen. Today we are diving into a topic that doesn't always get the attention it deserves, but is quietly shaping the future of AI development and deployment, and that is AI standards. To help us understand this space, I'm thrilled to welcome Owen Larder, head of Frontier Policy and Public affairs at Google DeepMind. Owen, thanks for joining us.
- [00:34]
B
Great to be here.
- [00:36]
A
Okay, so before we get into the meat and potatoes of the conversation, I always love to hear about how you wound up in the world of AI policy and also I guess how you ended up at Google DeepMind. So how did you specifically working on AI and AI standards?
- [00:50]
B
Yeah, thanks Greg. I mean it's sort of a story of luck and opportunism and good timing, I guess. My background is in politics and policy. I'd been doing that for a while and then ended up at Microsoft where I had 11 fantastic years doing various different things across across the UK and globally. I worked on government affairs issues, policy issues and then towards the end of that worked in the Office of Responsible AI, which I joined just before the large language model revolution kicked off. And then that happened and the.
- [01:18]
A
That's how you and I met when you were in that gig.
- [01:20]
B
That's right. That was how we originally got in touch back then. Yeah. But it's great now to be at Google DeepMind leading the policy team globally here. I'm about nine months in, I'm still early enough that the novelty of bumping into Nobel Laureates in the canteen hasn't quite worn off. And honestly the pace and breadth of research and application is remarkable. We're obviously very excited about how AI is already being used to do fantastic things. We were in India a couple of weeks back at the AI summit there and to see everyone from world class scientists using AlphaFold to develop new therapies, through to farmers using AI to improve crop yields. It's very exciting. Amazing innovation happening in the us, UK and elsewhere. But yeah, to segue into standards, I guess it's also very clear that this is powerful technology. It's going to have a very major impact which we're very intense to sort of think thoughtfully about. I think standards are going to play an incredibly important role in charting our courses. So they're not the only thing we do here, they're not the only way that we think about governance, but I think they can be a really important part of the framework and they have a really interesting history that I think we can learn a lot from as well.
- [02:23]
A
Yeah, and I should mention, you know, as you said, your Remit@Google DeepMind is way broader than standards. But you have offered to be our sort of subject matter experts to give a crash course in standards. Standards. Because everybody who is working on AI and governance in any kind of substantial way immediately discovers, whoa, the standards are going to be really important here. And so I appreciate you geeking out with us here on the podcast and giving us the crash course in standards. So let's start at the basics. I think a lot of our listeners have obviously heard the word standard, encountered the word standards, but would sort of struggle to define it in its formal sense as you encounter it in the technology industry and a technology policy. So how do you explain standards and why they matter to somebody who's never thought about them before?
- [03:10]
B
Yeah, yeah. So, I mean, there's nuance here. There are capital S standards, which are formalized standards that are developed through standard development organizations, and we'll talk about those. But there's also, in concept, you know, a standard is just an agreed upon and repeatable way of doing something useful, essentially. And this is something that we as a civilization have been doing for a long time. You had Carthaginian ships 2000 years ago that had numbers on the join so that people could assemble them in a standardized way. You had the Egyptians standardizing the qubit as a unit of measurement so that the pyramids all look nice and uniform. And then you have a long history from them through standardization of the width of a railroad track or dimensions of a shipping container. And then sort of latterly, really in the 20th century, this process for developing standards more formally has matured over time. And so you now have organizations like the International Standards Organization, for example, that will bring together a multi stakeholder group of experts on a particular issue and then work with them to sort of formalise an established domain of knowledge into a document that can set out this repeatable way of doing something. It's probably also worth talking about the way in which there are these formal standards, as I mentioned there, and then other useful things that point in the direction of standardized best practices, but are not quite the formalized standards. So guidance and best practices around how to do something you can think about in the AI context, the Frontier Model Forum, for example, where industry is coming together and sort of sharing early thinking about how we're doing things like testing for particular risks or mitigating them.
- [04:53]
A
Awesome. Okay, so you talked about the distinction between capital S standards and lowercase S standards, which I think is helpful. There's another Distinction of standards that I want to make clear, which is the difference between like technical interoperability standards, like the USB port standard or 5G telecommunication standards. And then on the other side there's process standards, which is, you know, the safety checklist that an airline has to go through before putting the plane back in the air. So I gave you two categories. What are the main categories of standards that we should understand? And can you give like concrete examples of each?
- [05:31]
B
Yeah, I actually think that's a fairly helpful delineation and they're both really important. So technical standards, as you mentioned, sort of often define how a technology ecosystem interoperates. And then you have the process standards that can support things like risk management. There is a long and rich history of both. So I think about the technical standards in the Internet space, which are really, really important in standardizing how computers talk to each other, quite frankly.
- [05:58]
A
Like Internet protocol is a standard Internet protocol.
- [06:01]
B
Hypertext transfer protocol, HTTP URL, these are all standardized ways of doing things.
- [06:09]
A
I guess it's just worth like, what problem does that solve in the Internet? It's like I. Let's just say I've never met you, you've never met me, but we both know that if we use Internet protocol, if we use hypertext, then our computers can talk to each other before we even meet. It's like it lowers the cost of.
- [06:27]
B
In the early days of the Internet, obviously the Internet was just computers being connected to each other with a pipe, but there was no standardized way for them to communicate. So unless I knew exactly where on your system a particular file is that I wanted to access, it didn't really work. So that's what HTTP basically standardizes the way that a browser can talk to a website. Ur a standardized way of giving an actual address to a website. So they're absolutely fundamental in terms of building the connected and useful Internet that we have. There was also then an issue of trust. Of course, I'm just about old enough to remember when people would tell you that there is absolutely no chance in heck that I'm going to put my credit card details into the Internet. That does not seem like a good idea, which it wasn't when you were just sort of communicating and sharing information around in plain text. But then you obviously develop this encryption layer, HTTP becomes ht and you're able to share encrypted information in a much more secure way. And that's basically what kicks off a whole boom around the Internet enabled economy.
- [07:27]
A
Yeah, so you talked about technical standards. Do you want to elaborate a Little bit on process standards.
- [07:33]
B
Yeah. And so then process standards would be used to do various different things in various different industries to, you know, test a piece of equipment to make sure it's working properly, test, you know, whether it's a car or a plane, make sure that it's adhering to a process that we stands to be safe.
- [07:52]
A
Yeah. And you use the car example. Right. And that can be everything. That could be like the crash test protocol of the car or that can be, you know, if you want to say that this steel is going to be strong enough to be in the car, it has to be heated to such and such a temperature, include such and such a purity in the right order and blah, blah, blah, blah. And at the outcome of that is like a steel that we can trust. Okay, so you've hinted at this already, but I wondered if you could sort of like put it in a little bit more formal categories, like what are the benefits of standards? Why do we care about standards? Why do they matter?
- [08:30]
B
So I think firstly, if we start to look at the AI space, you need standards to help technology talk to each other. You need interoperability standards for things like agents as we move into a more agentic economy. So we're trying to contribute to this as Google and Google DeepMind, we have the agent to agent standard and the universal commerce protocol, which is in many ways the sort of modern equivalent to, to the Internet standards that I mentioned before. You then also need these process standards to build trust in the technology so that people are happy to go and use it. And so I think there's more work that needs to be done there to develop standardized ways for testing systems, identifying risks that we care about and applying mitigations. Again, I think there's lots of good early work that is going on there, but I think it's easy to forget sometimes with all of the developments that we've seen in AI over the last few years, that we're still relatively early, I think in terms of developing this technology and applying it. So as the technology develops, we need to mature an ecosystem to develop these types of standards alongside it.
- [09:34]
A
Yeah, so you talked about interoperability, which is my thing is going to work with your thing. You talked about trust, which is I haven't done a huge due diligence of your technology, but if I know that it conforms to the standard, then I know I can rely upon it and I don't need to do the extended due diligence in many cases. What about the costs? Theoretically we could have a formalized capital S standard for every technology. What makes standardization challenging or less attractive?
- [10:15]
B
Yes. I think this is where we can sort of lean on history a little bit to learn some examples of when standardization has gone well and gone not so well.
- [10:22]
A
Yeah, the imperial system is when it's gone poorly. My gosh, we're still stuck here.
- [10:27]
B
No comment on that. I'm not going anywhere near that.
- [10:31]
A
Kilometers for the wind is all I'm going to say.
- [10:33]
B
Yeah, there you go. Something like that. The history of electricity. Actually, this is something we like to do in general at DeepMind. Sort of look back at previous periods of technological change and see what you can learn. I actually think that sort of late 19th century, early 20th century moment, including in relation to electricity, is incredibly illustrative. A lot of parallels to AI today. So electricity was another one of these technologies that was clearly going to be very impactful, but it was not totally obvious. The impacts were going to be sort of amorphous, general purpose technology that you couldn't really see. Quite scary as well. I mean, in the early years of playing with electricity, a lot of people setting themselves with the houses on fire
- [11:12]
A
being a real problem. Being an electrician. There's a lovely book called the Age of Edison, which is about the early decades of electricity. And yeah, being an electrician was not a safe job. It's a dangerous field. Early electricity is not inherently safe. We made it safe. We figured out what the right protocols, what the right standards, what the right technologies were. Some of them we enshrined in standards, some of them we enshrined in regulation. But, like. Yeah, I think you're right on. Keep going.
- [11:42]
B
Yeah, no, I think that's right. I mean, well, it's dangerous being electrician. Also dangerous being someone in a house with poorly insulated wiring for a long time. I mean, famously, you had President Benjamin Harrison at the end of the 19th century, who's too scared even to turn the lights on. He had his servants turn the lights on and then leave them on for the whole night because he was scared to be touching the light fixtures. So what happened that was obviously a major limiting factor on adoption of electrical appliances. And so you have this great story in the UK where this super team of scientists, William Thompson, later to be Lord Kelvin, and James Clark Maxwell, come together and they say, look, we need to come up with a standardized way of measuring electricity. You need to understand how you measure the resistance of wiring that you're going to be putting electricity into. And so. And they get together with a number of other scientists and they come up with these units that we all know well today, the ohm, the amp and the volt, which unlocked an enormous amount of value. So, firstly, if you're a power supplier, you now know how much current should I be putting into the Wyes. If you're in a product manufacturer, you know how to take the electricity off the grid and how you should be optimizing your equipment. It also then very importantly allowed you to build things like a fuse, which is to say, you know, if you, if current above a certain level goes through this fuse, it's going to break and we're not going to torch your microwave or whatever it may be. So I think that's a good example that we need to sort of build on around how standards are going to underpin interoperability and safety in the AI era. There's also, in the electricity space, good examples of what happens if you don't get this quite right. Again, having traveled recently and those of us that travel a lot, really annoying when you go to a new place and your laptop and your plug doesn't fit into the socket. Annoying. Creates an enormous amount of waste and also is highly problematic for manufacturers if they are looking to develop products and sell them across borders. So I think a good reminder that interoperable international standards, particularly when it comes to a technology like artificial intelligence that's going to be developed and used across borders, is something that we need to be prioritizing.
- [13:57]
A
And you're hinting at something which is maybe not the most important point, but, like, I still think is worth raising Sometimes, you know, standards can compete against one another. You have the US electrical standard that competes against, for example, the UK electrical standard. Like what, what a socket looks like, what a plug looks like, what the standard voltage is. All of that stuff is, is different. And you experience it when you travel and your, your plugs don't fit in the outlet. Now, sometimes that can be because, like, the, the two groups have differing opinions on what. Like, the UK is very proud of their socket. It's supposedly the least likely to catch fire. It supposedly, you know, has a bunch of other benefits. And so maybe, you know, they go a different direction because they think, I have a better thing. Maybe the Americans go a different direction because they're like, look, we've already installed this infrastructure. We're not going to swap out everything, you know, just to have this modest benefit in fire safety. But also standards could be a form of protectionism. You know, you could basically say, like, we want to have local champions. We don't want to be obliterated by a General Electric or a Westinghouse. And so we're going to come up with these alternative technical standards and our local companies are going to be good at them and foreign companies are not going to be good at them. And it's actually not going to benefit consumers, but it is going to benefit our local industry. So sometimes you see this phenomenon where standards are a source of protectionism.
- [15:26]
B
That's exactly right. And I think you can see this sold in a number of different ways. Sometimes the market just wins out and the most popular standard is irresistible. There are also examples of where exactly what you just described happens, and I think that could be problematic. And again, we think about AI itself, which is just an inherently international technology which is developed and used across borders. It's going to be very generally applicable across, you know, pretty much every aspect of every society. I think sort of locking things down and balkanizing standards is going to be something that was not helpful in terms of unlocking the benefit of AI.
- [16:04]
A
So now let's talk like one of the other downsides of standards is sometimes the process is slow. So can you kind of walk us through soup to nuts? How do you make a capital S standard?
- [16:22]
B
So capital S standard is typically where you have an established standards development organization. So let's say the International Standards Organization.
- [16:31]
A
And what the heck is that? What is the ISO? What is the standards organization?
- [16:36]
B
This is. Yeah, I mean there are a number of different standards organizations out there, but the ISO is certainly a very well known one. It brings together experts from across various different parts of society in a global way. So this is an international organization as, as suggested by the title. The expert groups that they convene do various different things. So they have working groups that in a particular area will say, do we need a new standard to be developed here? Then there is an opportunity to propose a new standard. A lot of time a proposal will be struck down. The suggestion will be that this isn't needed or it's duplicative. But then if a gap, an opportunity is identified, then you have these multi stakeholder working groups that in quite a deliberate and structured process will review a proposal, carry it forward, put it out for opportunities to comment, vote on it, and then sort of either based on consensus or sort of broadly based on a sort of consensus based approach, you will eventually, over the course of the 12 to 24 months, come up with a new standard in a new space. Now it is interesting when you think about the speed with which AI is developing because this is a technology that is, has covered an enormous amount of ground over the next couple of years and capabilities continue to improve. I think it's very likely that what AI can do and the way that it's going to be applied in the world will change in the coming years. It's also quite a new technology. And if you think about what standards organizations typically do, they bring together areas of consensus or consolidate your expertise and then formalize it into processes that people can rally around. So in the AI space, I think we're going to have to try and find this pipeline essentially of sort of research about the technology to properly understand it. Let's say it's in relation to testing methods and then take that, formalize it into sort of standardized best practices and then put it into these international standards organizations. That's going to have to be sort of an ongoing process over time, I think, so that we can keep pace with the development of the technology. I do think it does offer some flexibility when you're thinking about governance frameworks broadly, because this is complicated technology, you're probably going to need some kind of deliberative process around what the appropriate way of governing it is so that you're realizing the benefits and mitigating the risks. I think standards will play one part of that.
- [18:59]
A
Yeah. And there's a term of art that I think is important to grasp here, which is the standard essential patent. And basically, you know, some of these standards, these technology standards cost money to develop. And some company might say, I developed the right technology, we should make my technology the standard. Well, once it's the official standard, right, kind of sort of, everybody has to use it. So you can imagine it would be very lucrative for a company to earn this spot, the spot where their intellectual property is written into the standard. And one of the interesting things that happens is like the negotiation between a company and the group at the standards making body where they'll say like, well, if you're going to like, like for, let me give you an example of Qualcomm. Qualcomm invented a lot of the standards for the technology at the heart of gsm, which is super important for every single cell phone. So once they're in the standard, you know, what if they say that the cost to license our patent is a million dollars per phone, right? Well, everybody's gonna be like, wow, we really regret making that one the standard. And so you actually have a negotiation where a company will say, well, what if I only charge this much? Will you make my IP the standard? What if I Only charge this much, will you make my IP the standard? And so there's kind of an interesting intersection of standards and intellectual property that come together now, just so people can understand like Google's role in this story, whether AI related or not. Can you give an example of a standard where Google DeepMind was influential in the process?
- [20:50]
B
Yeah, so this is something that we see as almost a responsibility to be able to contribute to these discussions on standardization. I could give you some examples both on the technical side and then on the process side at your discretion.
- [21:04]
A
Yeah, yeah.
- [21:05]
B
So mentioned previously these sort of agent standards that I think are going to be really important to interoperability of agents. So we have our agent to agent standards standard, which is basically a way of connecting agents so that they can talk to each other. Again, real parallels to the early Internet days. If you want to connect two agents at the moment, it's actually slightly tricky. They have to be either sort of both working within the same walled software garden or you're going to write some bespoke code to connect the two. What the agent to agent standard does is essentially have like a standardized clipboard of information that two agents will share with each other as they come together. You know, this is the ID of the agent, this is the type of data it takes, this is the capabilities of the agent and what it's trying to do. Sort of like a handshake that makes it very easy for them to interact. Universal Commerce Protocol is something similar, but relates to how an agent would talk to a website. Again at the moment, often what you have is sort of an agent will click on a website to try and buy something at the instructions of, you know, human that is tasked it to do that. And it will have to sort of grasp around and sort of try and work out what does this website even do? Is there a way I can make a payment? Where do I click to do so? Again, what you can do with the UCP is it's a sort of standardized way of sharing the information about the website, what it does, if you know information about payments. So this is something that we feel.
- [22:32]
A
These are still very much being debated. We're probably not going to have a formalized standard on this today or this month, right?
- [22:39]
B
Yeah, exactly. So these are out there being used now. We've done this in partnership with various other organizations as well.
- [22:46]
A
Oh, interesting. So you have issued sort of like a thing that you say this is the Google standard for now, Google is always going to use this and we also recommend that the world use this. Okay, that's right.
- [23:00]
B
We've contributed this as an open source standard that we think is useful to solve this particular problem.
- [23:05]
A
Got it. And when you do that, are you like going to the ISO and saying, hey, make our standard the universal standard or are you primarily pursuing the open source path or how does that work?
- [23:15]
B
No, we haven't gone to the ISO on this one. This is an open source standard that we have developed and put it out there, working with various different partners across the world to do that. So that would be an example of how we lean into the sort of technical standards. There is then also the process related standards that we mentioned. This is something where I think we as an industry are learning an enormous amount about the capabilities of AI and really good effective and repeatable methods for actually testing, understanding how the particular model or system works, thinking through the potential risks that it may pose when deployed in the real world and then, and then coming up with mitigations. I think this is something where an enormous amount and sort of ongoing research is needed as the technology continues to develop. So we're trying to lean into organizations that are pulling together knowledge of these types of things. I mentioned the Frontier Model Forum, this is an industry group that we helped set up that the other leading frontier AI companies are in really useful to be able to sort of exchange knowledge and understanding of these developing practices. You have other organizations like the US center for AI Standards and Innovations, the UK Security Institute, I think also really helpful as we all learn about this technology and how to test and address risks and so we try and contribute to conversations that they are developing as well. You will have seen that the USKC has launched this Agent Standards initiative which I think is a really helpful thing to be doing and we'll look forward to contributing to that as well.
- [24:47]
A
So you mentioned ISO, UK AI Security Institute, the uskc, which is part of nist, the National Institute to Standards and Technology. Are there any other bodies that are big in the AI standards space that folks should be aware of what they're up to? You're probably not going to be able to name everybody who's relevant. This is it.
- [25:07]
B
And you'll always leave someone out and run the risk of upsetting people. I think the point here is that this is a sort of broad and nuanced ecosystem that has a large number of players, some sort of international formal organizations that you mentioned. Actually you have a lot of sort of just industry specific coalitions that come together to solve specific problems. So whatever your sort of issue or interest, there is probably a Standards organization out there for you somewhere. But I think we sort of covered a lot of the well known ones in the conversation so far.
- [25:38]
A
So I think one thing that was prominent in AI policy conversations a couple of years ago was the NIST's AI risk management Framework. Obviously, even though it's coming out of the National Institute of Standards and Technology, it is a framework, not a standard. Um, but can you explain like what the, the risk Management framework is? And are, are there any other major AI standards or frameworks that folks should know about?
- [26:06]
B
Yeah. So again, this has been a developing conversation over the last couple of years. So the NIST AI RMF as you mentioned, came out, I guess a, a few years ago and gave some guidance around how you could test systems for various different capabilities, various different attributes and, and how you would mitigate them. I think you've seen a lot of other developments and suggestions since then. So you have guidance coming out from Casey, which is also part of nist. Lots of good work coming out from the AC as mentioned. But there are lots of standards that are also already in development. I mentioned some of the ones in relation to agents. You have standards that are coming out of the International standards organization, like ISO 42100 on AI management framework for AI. So there's been a lot of sort of early offerings, but I think more needed as forward.
- [26:56]
A
Great. Okay, now I want to talk about the intersection between standards and regulation. Folks who've been listening to this podcast for a while will know that I've, I've pointed out that the EU AI act has a very prominent role for standards. Some of the regulations basically take the form of thou shalt follow the standards. So can you sort of like walk us through maybe some of the standards that really matter in the EU AI act, whether they have been developed or whether they're still in development. And just help us think through the distinction between voluntary standards and mandatory standards more broadly.
- [27:31]
B
Yeah, and I think that's a helpful starting point. So standards are distinct from regulation. A lot of standards are voluntary as they're being developed. You've seen this trend in the regulatory discussion over the last few years to sort of cross reference standards in a way that I think might be helpful over time so that standards can continue to develop and keep pace with the technology and regulation ends up being more durable. The AI act, as you mentioned, cross references what are known as harmonized standards. So European standards that are going to be developed to underpin the various different layers of the AI regulation, high risk systems, and then also at the model layer, you see other regulation that is being developed around the world also referencing standards. So SB53 in California, for example, a frontier regulation, talks about standards and the need for organizations to talk about which standards they're using. I think this is something that we will sort of see be a part of the regulatory discussion as it moves forward. I think what it also means is that work needs to be done to develop these standards that are being cross referenced. So you're seeing quite a lot of efforts going on in various different parts of the world, the EU and internationally, to develop these standards that will then be cross referenced by the regulation going forward. So it's a little bit funny because some of this regulation is almost at, not necessarily placeholder, but sort of gesturing to things that have not yet been developed, which I think puts all the more of a premium on getting those standards right, making sure that they're going to sort of be supportive of innovation whilst also addressing the risks.
- [29:14]
A
Yeah. And I think to share a little bit of inside baseball in the EU AI Act, a lot of this said, thou shalt follow the standard standard which we will come out with very soon, you know, and then the standards making organization that is in the European Union, very prominent senilac talking with some of those folks, they're like, oh my gosh, how could we possibly develop a standard under the time constraints that you've given us? And one thing that's interesting about the European Union and this comes up in the AI act is they have the Vienna and Frankfurt treaties which sort of gives a default preference to prominent international standards bodies like the ISO. They don't want the EU sort of reinventing the wheel on, on standards bodies and that, that comes into conflict sometimes, I would say with those who are interested in potential opportunities for European protectionism or technological isolationism, you know, viewing local standards development as an opportunity to, to build a more domestic technology industry. So it's, it's, it's really interesting how like sort of different political factions, different political incentive structures intersect with the standards world. So we've already talked about the generic costs and benefits of standards, but specifically when it comes to AI and adoption, like how can standards help or harm AI adoption?
- [30:40]
B
Yeah. So I think go a little bit back to what we talked about previously. I think you're going to need to develop these technical standards to have the interoperability unlock for the AI world, for the agentic economy in the same way that we had there for the Internet era. A lot that needs to be done there. There is then also this issue of trust with what is still a relatively new technology. And I think until you have a good process for developing those process standards and reassuring, whether it's individuals or companies, that they can use this technology and be compliant with regulation and also stay on the right side of governance challenges, I don't think you're going to fully realize the benefits of, of a. Yeah.
- [31:22]
A
And so the, the competition within the standards bodies to me seems to increasingly reflect like geopolitical competition. And I think this shows in the Trump administration's AI Action Plan, which directs the Department of Commerce and the Department of State to, quote, position in international diplomatic and standard setting bodies to visit vigorously advocate for international AI governance, appropriate approaches that promote innovation, reflect American values and counter authoritarian influence, end quote. So what do you think listeners should understand about the intersection of standards and geopolitics?
- [32:00]
B
Yeah, I think you have a few fundamental principles that we should be guided by here so that standards should be expert led. Particularly when it comes to a technology that is as complex and nascent as AI, you need the right experts around the table sharing perspectives. A lot of those will be in industry, a lot of those will be outside of industry as well. You need good representation. You also need to make sure that these standards are being developed, I think in an international way that works across borders. I keep coming back to this, but I think it is a fundamental characteristic of AI technology that it is going to be developed and used across borders. I don't think it's in anyone's interest to have a sort of balkanized approach to the technology. And I think standards should also be open and facilitate interoperability as well. We've learned this from previous eras of technology. So I think leaning on those principles, I think working within the sort of existing structures, I think that's going to be the right way to go.
- [33:00]
A
Yeah. Let me add one thing on the geopolitics of it and connect it to something I said earlier on standard essential patents is it certainly comes up when the US tries to impose its will using export controls. So when there were export controls on ZTE and Huawei in China, China, sorry, the United States imposed such export controls against those companies in China in 2018 and 2019. Part of the reason they were so damaging is that they were restricting the export specifically of American chips that were part of the standard essential patent. So we said you can't buy chips from Qualcomm, which in the case of ZTE basically meant you can't build cell phones or at least you can't build cell phones without, without violating Qualcomm's intellectual property illegally, which may have been willing to do in China, but they certainly weren't willing to do, you know, abroad around the world. And so that, you know, is, is a source of potential geopolitical leverage. And I think more and more countries are conscious of, of this as a decision making criteria as they approach the sort of standards conversation. But let's, let's bring it back to, to Google DeepMind. So you've talked a little bit about how Google has influenced the, the standards of different practices and is standards. Can you give me an example of the reverse in which something going on in the standards development world shaped the behavior of Google or Google DeepMind?
- [34:37]
B
Yeah, I mean to go back to some of the examples we gave before, Google is an organization that is literally built on the Internet standards that were developed through the 80s and 90s. I think the same thing is going to happen in the AI era as well. I think as we develop these standards and contribute to these standards bodies, they're going to be a really important, important part of how we do our work across the company.
- [35:01]
A
Cool. And you're an important and busy guy, so I know we're going to lose you in not too long, but I want to conclude with your sort of advice and recommendations to policymakers and policymakers from all over the world. I'm proud to say. Listen to this podcast. So what do you think is some areas where progress is most urgently needed and what would you, you recommend?
- [35:24]
B
Yeah, so I think this is an evolving space where we all need to learn together. I think that is just going to be a defining characteristic of the technology that is going to continue to move quickly, going to be complicated. So I think we need to invest in some of this pre standardization work that I was talking about, really try and develop this pipeline of sort of safety research and understanding understanding into guidance and then be flexible and iterative and build that into formalized standards over time once we build the knowledge needed to do that. I also think just being very much led by the sort of the science and the expertise around people that understand this technology well is going to be an important part of getting that pipeline and that sort of durable infrastructure. Right?
- [36:13]
A
Yeah. And are there any standards initiatives, you know, going on in the United States government that you want to call out or around the world that you want to call out as good work that deserves support?
- [36:26]
B
I think there's great work going on in a number of places. I mean, I maybe will call it out just because it's it's very live and recent, but the, the effort from the U.S. casey, around this agentic standards initiative, I think that's going to be an important one to, to contribute to. I think the work of the International Standards Network of AI Security Institutes is going to be a really important sort of institutional bedrock of this work going forward.
- [36:51]
A
Yeah. And I think that's noteworthy that you know that that group of international AI Safety and Security Institutes is not the entire Earth. Right. But it is a group of influential American allies and in America, of course, itself. And it's one of those situations where, where they could have global influence, but only if they do really high quality work that the world wants to imitate and adopt. And so it's really important to invest in this and get it right because quality is a key input of the opportunity to have leadership and to demonstrate leadership.
- [37:31]
B
I think just, just a. Yeah, please double, double click on that and endorse it. I agree. I think the high quality bar is really, really important. And I think that network work is very international. I think it has good representation from around the world. So I think it is an important institution for sure.
- [37:48]
A
Awesome. Well, Owen, I've known you and I've learned a lot about this topic from you over the past few years. I'm now grateful that you're willing to share it with our audience here on the AI Policy Podcast. As I said before, standards is only one part of your enormous policy portfolio. So I'm sure we'll find an excuse to have you back on the podcast at some point or another. So thanks much for coming on.
- [38:13]
B
That would be great, Greg. Yeah, it was great to great to discuss this and thanks for having me on.
- [38:18]
A
All right, thanks for listening to this episode of the AI Policy Podcast. If you like what you heard, there's an easy way for you to help us. Please give us a five star review on your favorite podcast platform and subscribe and tell your friends. It really helps when you spread the word. This podcast was produced by Sarah Baker, Sadie McCullough and Matt Mann. See you next time.