Industry Innovations with Google Cloud

A conversation with Google Cloud’s Patricia Florissi about the digital dichotomies of the decade and the role of government.

Alistair Croll: [00:00:00] Hi, and welcome to another interview with one of our partners for FWD50 of our Industry Innovations series. Today, I am thrilled to welcome Dr. Patricia Florissis . She is the technical director in the office of the CTO at Google. She’s been working in the technology industry for decades and has fascinating thoughts on the trade-offs that we have to make as we adopt technology and make it part of our everyday lives.

Welcome, Patricia. 

Patricia Florissi: Thank you. Alistair. And it’s a pleasure to be here today. Humbled by it. 

Alistair Croll: It’s great to have you. So when we were discussing this conversation earlier, you mentioned that technology has a lot of benefits, but a lot of drawbacks, and these really are trade-offs that we need to think about as humans. Briefly before we get into each one. Can you talk about what those trade-offs are? 

Patricia Florissi: Yes. I will name just three there are many more but three that come to top of my mind. [00:01:00] One is privacy versus unbiased fair AI. The second one is sustainability versus pervasive ubiquitous AI. And the third one is sovereignty versus scale and speed of innovation.

Alistair Croll: So those don’t sound like trade-offs on their surface. So maybe you can give me a little more detail about the first one you mentioned. 

Patricia Florissi: Okay, so let’s take a deep dive on privacy versus unbiased AI. We want data privacy, but also AI to be as free of unfair bias as possible, which requires information such subgroups and demographics in order to accomplish “unbiasedness”. Can we achieve both?

Alistair Croll: The example that comes to mind immediately for me is that when I tell Google maps my location and how I’m going to get to a destination, whether I’m walking or driving, these are all information that is essentially private information about myself, particularly if I’ve added things like I [00:02:00] like sushi, but I don’t like steak then it can make really good recommendations about “here’s, how to get there and which restaurants to stop at on the way”, which feels very utilitarian and very useful—but I had to disclose something about myself. Usefulness increases with context and I was reading Stewart Brand’s amazing line, “the right information at the right time just changes our lives.” It seems like that’s a very ominous or important sentiment when it comes to wanting AI to change our lives for the better. 

Patricia Florissi: You are very, very right. On one hand, we are committed to building technology that works for absolutely everyone, right? So we can be developing advanced technology that is aligned with AI principles, especially Google AI principles and human rights in areas such as privacy, unintended bias and inclusion. And these includes not only respecting the legal and regulatory requirements, but also considering things like social norms and typical individual [00:03:00] expectations. On the other hand it’s necessary to have access to appropriate data in order to test, validate, and train models. And we recognize that we do not want in many ways to compromise the reach of the data that is being analyzed, and instead focus on promoting data inclusion and data equity in order to accomplish our fairness commitments. And this is an ongoing area of research, to be honest with you in the ML community, with significant room for growth. 

Alistair Croll: Well, I remember Susan Etlinger—who’s one of our speakers this year —first introduced me to Word2Vec, which is the big Corpus of name analogies. And unfortunately when you look at that data, in the same way that it would say “Paris is to France as Tokyo is to Japan,” it would say, “man is to doctor as woman is to nurse,” which is obviously very unfortunate. But at the same time that there were efforts underway to identify these inherent biases.

And to me what’s [00:04:00] most interesting is that many of the things that we’re now, quite rightfully, outraged about always existed as prejudices, but we weren’t shown them. And it seems to me like the first thing AI did was hold up a giant mirror because it’s trained on us and now we can see where we need to intervene.

And a great example of this is the other day I did a Google search for the new James Bond movie called No Time To Die. And I typed into Google “where to see no time to die”. And the result I got was not movie listings. It was a suicide hotline because I typed in the word ‘time to die,’ and Google went, “Hey, the correct response to someone asking ‘is it time to die?’ is to show them a suicide prevention hotline.”

So it’s not just about AI on the training side. But it’s also about context of the search and what’s the right thing to do on the query and display side. It seems like that whole pipeline requires thoughtful reflection of the [00:05:00] trade-offs. Is it clearly a person who’s feeling okay and can find another way to Google a movie time, whereas someone who does need help probably should have had that intervention. And so to me, it’s fascinating how we navigate those trade-offs between disclosure, privacy, and then what the data is telling us. 

Patricia Florissi: Again, bringing up an important point because AI is not starting from scratch as AI is not a self-isolated application. On the contrary, AI technology is integrated into a workflow and environment where it is making decisions based on what it knows. And those decisions impact people in a very, very different way than creating software, like an e-commerce platform.

And what it knows is determined by people who are at the center of, and at each stage of machine learning development. And throughout the course of history, what we know is that people also have a history of making decisions that are not in line with the needs [00:06:00] of everyone.

So it is because of this that AI systems are a reflection of what exists in society—which was one point that you made—and without good practices, AI may replicate existing issues or bias and amplify them at unprecedented levels. Human rights, and the development of AI principles, along with an approach to operationalize those principles in our organization, can serve as a starting point in understanding and addressing disputes and address cases as you said, and I believe that there are many practices and steps in making AI as free of unfair bias as possible. And one of which is taking a multi-stakeholder collaboration is not just one side, but one that will allow for several sectors to to participate: the civil society, government, technology providers, academia, sectorial, or industry [00:07:00] experts. It’s a collective effort.

Alistair Croll: So Patricia obviously there’s this trade off between privacy and effectiveness. The more you disclose, the more effectively a system can tailor its results to you. How are companies—and Google in particular— finding a way to balance that and, and share as much as possible with the tool that’s working in our service without leaking that data elsewhere, where it might be used in bad ways?

Patricia Florissi: Oh, thank you for asking this question because I am very excited about everything that Google is doing. In 2018, we published Google’s AI principles to actually help guide ethical development and use of AI. These are really broad corporate guiding standards, which are backed by robust internal processes to ensure that they are implemented application by application to individual businesses and product decisions.

So let me give you a high level of those AI principles: 

– AI should be socially beneficial.

– [00:08:00] AI should not create or reinforce unfair bias. 

– AI should be built and tested for safety.

– AI should be accountable to people. 

-AI should incorporate privacy design principles and so on and so forth.

So as you saw, fairness is one of our core AI principles and we believe that well-designed ML systems can help to overcome unfair biases in human decision making and deliver more fair—and I would say equitable—outcomes for everyone. Our work spans the gamut. So to give you a few of them from a research area, we invest in research in areas like measuring bias and building fairness constraints into models; from open source tools like facets, and what-if tools, to help visualize representation and biases. We also provide online tutorials and community support. So we [00:09:00] really have a full gamut of angles that we actually work on to ensure that we are making the right trade offs here. 

Alistair Croll: And those are high up in the organization kind of trickle through Google compute and all the other platforms and tools you’re building. Is that right?

Patricia Florissi: Everywhere. 

Alistair Croll: So you mentioned there are three, trade-offs obviously the trade off between privacy and the utility or effectiveness of machine learning and algorithms is one trade-off. What about the other two? 

Patricia Florissi: So the second one is around sustainability versus pervasive AI. We want sustainability, but we also have a need to analyze huge amounts of data, unprecedented amounts of data for how so how can we leverage AI to the fullest potential without an impact on carbon emission. So as AI becomes ever more helpful to our personal and professional lives and the cooperation’s lean on a path of continuous expansion, as you probably have seen on the application and adoption of [00:10:00] AI the one aspect we often overlook is the fact that AI can be very expensive from a computational perspective, more those get trained or hundreds of thousands, or even millions of data set points that need to be collected and computed on and train them all. Those are not a single time event where I’m other strain one, one center who hint over a longer period of time as new data is gathered, then the world changes and evolves the same. All those need to be continuously retrained to learn and adapt to avoid any model drift and because of latency reasons more and more computing is being deployed in the at the edge outside of data centers, outside of clouds, incurring additional capital cost to users as they buy their own devices and to enterprise as they deploy something as IOT. And this is where the trade off is born. On one hand, AI is expected to be ubiquitous and coherent, [00:11:00] and we rely on AI more and more for all day lives. And on the other hand, sustainability has become another fundamental digital human rights. So how can we balance both. 

Alistair Croll: Yeah, it seems to me like there’s an interesting philosophical thing going on here, which is that, you know, there’s a sort of crystallized knowledge, which we would normally think of as wisdom, the stuff that you’re taught, the stuff that you learn over time and file away for quick use. And then there’s this sort of fluid intelligence built up on first principles. And in the same way, you know, the moment the model is computed it doesn’t know anything else. If you ask it to predict the future, it’s always doing it based on the past. And so the rate of change of the past tells us how frequently we need to update the model. But if the future, if the past keeps changing faster and faster then like the best before date of wisdom is getting shorter and shorter, which implies a greater carbon footprint. So what kinds of things can we do to reduce that? 

Patricia Florissi: Yeah, that’s a wonderful, wonderful question because they, [00:12:00] environmental impact of ML models depends actually on the amount of energy that is actually used than the source of that energy. And that lies the, the secret because we are constantly trying to innovate to improve ML efficiency so that it will consume less energy, which is good for sustainabiliy and costs, but we have to look at where they’re at. The energy is coming from at the Google cloud and Google in general, Google’s operations are carbon neutral. So we have no net carbon footprint for, for training and serving ML models at our data centers. Now it is important to understand betwee the difference between carbon neutral and neutral, and carbon-free. In carbon neutral we buy enough renewable power to counter attack emissions from dirty energy sources that that might be used to run a data center at any given point in time. But Google [00:13:00] has made a commitment that by 2030, our operations will be carbon-free. All the time everywhere, which means that all of our ML models will be trained and served using entirely carbon free energy. Now, let me give you a little number here to put the number, really capabilities. We estimate that the training for Lambda, a large model with 2.6 billion parameters had sealed to equivalent emission of about 96.4 metric tones. And to give you an idea, this is as much carbon as a one way plane trip from San Francisco to New York to train one single model. So we believe that the benefits from machine learning from more efficient and fast retrieval of auto climate changes solutions are worth the energy use to train ML models but we also share our findings [00:14:00] through publication and open source code so they can be applied across the field of ML research.

Alistair Croll: Effectively there is roughly zero atoms of carbon give or take a bunch, but you know, roughly zero carbons of atom in the atmosphere that are the result of Google.

Patricia Florissi: Precisely. So you don’t have to feel guilty when you do a search or when you search too. 

Alistair Croll: Are you starting to see procurement requirements that, that stipulate carbon zero or carbon neutral companies? 

Patricia Florissi: I don’t know if there are regulations per se, but we are not only promoting carbon free users and making a commitment by 2030, but we are also enabling our customers to make a choice of where they want to store their data and where they want to run their computations between the cabin and between locations that differ on carbon emissions. So we wanted to create incentives and a practical way for our [00:15:00] customers to actually move workloads when possible to regions that are more carbon-free 

Alistair Croll: And I guess workload movement is really important too, because as you see spiky workloads, some parts of the world are in high demand. Let’s say it’s warmer there, or it’s daytime there. And as a result, air conditioning is running or it’s cold or there, and heating is running. Being able to move workloads around with a relatively small difference in, in latency, because you know, if you’re training a model, It’s happening and then you get the results. It’s not like real-time communication. Being able to move a workload around the world geographically so that you’re using places that otherwise have idle capacity rather than trying to store it in batteries is a good reason to use elastic computing of some sort, the, the portability of a workload allows you to more effectively keep the level of electrical consumption flat rather than dealing with spikes.

Patricia Florissi: And that’s where customers benefit from the consistent system of development, building, operating, and managing their environment that they get with [00:16:00] Clouds and what we call open clouds approach that are based on open source and based on on an ecosystem. And that provides agility and flexibility through profitability.

Alistair Croll: Yeah. Yeah, the idea of my, my container, you know, Kubernetes and Docker sort of flying around the world, following the sun is kind of a cool idea. I was actually remarking on this the other day. My girlfriend pushed a button on her computer that made something change on a projector elsewhere in our house. But the act of pushing that button went to a data center using her software that sent a message to another server and then back to this, you know, other projected that was 10 feet away. And my brain kind of exploded it, thinking of the stack of resources, not just, you know, mining the Silicon to make the chips and the computers and ship it to us. But the mechanics of pressing the mouse button and sending a message to a data server somewhere in the cloud that changed some, some memory registers that pushed another message out. Like it’s almost [00:17:00] unthinkable how we have come to depend on this very complex stack to do something that, you know, our ancestors would have just yelled across the room, but we’re pressing a button and it has to travel halfway around the world on this incredible tech stack to send that same message, which I guess brings me to your third dichotomy.

We talked about the trade off between, you know, disclosing more information between wanting privacy versus the more you say, the more effective the algorithm is. And we talked about, Hey, we really want fresh insights from recently trained models at our fingertips all the time, without a large carbon footprint. But the third trade off that you really talked about, and this is the one that was more opaque to me when you first explained it was a scale versus sovereignty. So can you start by explaining what you mean by sovereignty? 

Patricia Florissi: I think that there’s a very, very important question here because cloud computing on one hand is globally recognized as the single most efficient, effective and [00:18:00] scalable path to digital transformation and to drive a value creation. As you know, it has been a critical catalyst for the growth, allowing private organizations and government to support consumers and citizens alike to deliver services quickly without prohibitive capital investment. And that enables unprecedented scale, but at the same time organizations, and I would say, particularly in Europe, they are leading an effort around the, and this includes both public and private sectors, provide that they’re looking for a provider to deliver Cloud on their terms. One that means meets their requirements from a security, privacy, and digital sovereignty without compromising on functionality or innovation. So let us take a deep dive on this concept of not just sovereignity but digital sovereighnity, which is even an arrow piece of the of the concept. At Google Cloud as a result of our conversations with [00:19:00] customers and policymakers, we developed a strategy based on three distinct pillars. Digital sovereignty is the union, if you will, of data sovereignty, operations sovereignty and software sovereignty. So let me talk about each one of the three at a time. So to give you a little bit context or a little bit of context, so let’s first talk about data sovereignty. What that means: this means that customers want and need and should have full control over their data, including storage and the management encryption keys. So for examples of customer controls that we provide today at Google Cloud include that the customers can store and manage their own encryption keys outside of the Cloud, as well as inside giving customers the power to only grant access to the keys based on the tail access justification. So, which means that even Google Cloud operators cannot [00:20:00] access the keys without offering a justification of why they are doing so and protecting data in use as well. 

So we have encryption of data in use. So that is the data piece. Giving costumers control over their data. It’s their data belongs to them. The second one is around operational sovereignty, which gives our customers total visibility and control to go back to the example of the process that your girlfriend went through. We believe that costumers need and should have visibility and control over Google’s operations. And we provide transparency, respectability, you brought the issue of audit ability before and security operating in a zero trust environment. And last but not least. And this is really the third pillar of the strategy sovereignty. This is also often also referred to as [00:21:00] survivability and the means the protection against unforeseen catastrophic events or what is called the black Swan. But here is giving costumers or providing customers the ability to run and move cloud workloads without being locked into a particular provider, which can also lead to better sustainability as we discussed including extraordinary circumstances, but the software sovereignty provides really costumers with assurance that they can not, they can control their availability of their workloads and run them wherever they want without being independent or lock them in a single cloud product. 

Alistair Croll: I’m reminded of a line of Henry Ford, who, when he was making the automobile said, you can have any color you want as long as it’s black. The key to scale in mass production is sameness. It’s standardizing. And so when you talk about scale, what you’re really saying is we can [00:22:00] only scale this if you all agree to follow the same rules, otherwise we are a bespoke and we can’t achieve the economies of scale that you get from mass production.And so if I summarize these properly, you’re talking about a trade off between the right to privacy versus the, if desired effectiveness of machine learning. You’re talking about ubiquity and freshness of machine learning versus its total cost, which includes things like the, the carbon footprint, which is often overlooked as an externality. And then you’re talking about the scale versus customization of the trade-off of the machine learning or compute. And I guess that has become much simpler in the cloud world thanks to the standardization of containers so that now you can create a workload you can make it portable, you can move it for on-prem to in the public. So let’s, I mean, this is a great philosophical conversation. And I think if we really want to get philosophical, we’re actually talking about the difference between individualism and collectivism, that [00:23:00] privacy you know, cost and footprint and impact, customization. These are all personal things but at the same time, you know, effectiveness for all ubiquity of compute and AI everywhere large scale compute that’s effective and cheap and can follow the sun. These are all collective benefits and there’s probably nowhere else that we run into this trade off between individuality and collectivism as there is in government, like this is literally democracy is collectivism constitutions or the rights of the individual, you know, there’s this inherent trade-off here.

S o how do you think the lessons that you’ve learned working in Google Cloud can be applied to the balance that governments need to make in both respecting individual privacy, individual rights, individual choice, and then giving people sort of standardized collective value at scale, where that’s useful for them as well.

Patricia Florissi: And [00:24:00] so I will offer a one view, there are many, but we believe that government governance should promote an interoperable approach to standards, to AI standards, to sovereignty standards, and also. So let’s, double-click on AI. For example, AI regulatory frameworks and technical standards need to operate not only domestically, but also across borders as in similar domains. For example, cybersecurity regulations should also allow for some level of flexibility for the standards approach to be chosen to suit the particular context. And I would like to maybe share with you something that happened in the US, just in June these year that the the current administration lounge, what is called the national artificial intelligence research results task force and the task force members will help develop a roadmap to democratize [00:25:00] access to research tools that will promote AI innovation and few economic prosperity. So now you have here the ability to give users access to new AI capabilities, including an excellent, including by expanding access to data such as government data that has been historically difficult to access, but that would enable further researching two key areas, such as bias and fairness.

So the government becomes an enabler of making AI technology, reachable of making data reachable and at the same time, providing interoperability. And this is because sharing certain types of government data and providing powerful AI tools raises important questions about how to protect people’s privacy and rights again but there are a number of steps that governments can take to protect privacy and civil rights, [00:26:00] including for example, ensuring the appropriate expertise among staff, reviewers and users and evaluating proposals for privacy and civil rights protections. So I believe this is a great step. 

Alistair Croll: So if I re if I want to recap this, I’ll make sure I’m understanding your properly, the regulatory from an AI perspective, consistent global regulations on AI can balance privacy with the effectiveness rafter on the ubiquitous computing kind of thing, which is like what I, what I need my AI everywhere, but also compute everywhere IOT, everywhere wifi ever, or whatever we’re looking at focusing on carbon neutral, digital footprint, which is broader than just like, is your stuff coming from a dam, but like a carbon neutral digital footprint for everyone will allow us to have ubiquitous computing and fresh data models that are less biased while not ignoring the full cost of that footprint, including the climate footprint. But then from a sovereignty point of view, we’re talking about standard [00:27:00] portable compute models that allow you to export and move data around, ensure resiliency. So you can have your own backups and computing your own term and implement governance models for that software while still taking advantage of these incredible economies of scale and sort of follow the sun elasticity around power consumption and computing ability. 

Patricia Florissi: Brilliant. And I would just add on the last one on sovereignty the importance of giving control to data users, making sure that there is transparency and inspectabilityand audiditibility to have everything that happens because that really fosters thrust on the cloud provider. 

Alistair Croll: And I think trust is something that definitely we need through our government more and more. So one final question, and obviously we can carry this conversation on for a long time. I think there’s a lot of philosophy here about collectivism and individualism, sort of behind this. Where do you see global standards for cloud computing going in the coming years around [00:28:00] regulation, governance, privacy, carbon neutral, like what do you think we’re going to get to a point where there are standards that make it possible to move workloads and to move data and to move processes without a lot of rewriting, without a lot of lock-in because that’s always been the concern is that if you move to a cloud provider and you take advantage of their services, to their full extent, it becomes very hard to move to another cloud provider. And so there’s theory of portability. It’s like a thousand tiny cuts that you have now become addicted to a particular platform and its API APIs and its function calls, and it becomes much less portable.

Where do you see standardization around cloud computing and portability going in terms of how governments can then adopt those technologies without giving up their sovereignty? 

Patricia Florissi: I think that is maybe the the holy grail. And I think that we believe in this concept of open cloud and it is rooted really in three principles. The first one is open [00:29:00] source. How can we actually make sure that whatever we build is based on open soft and people have choice and people have a way of influencing the shape of the technology that is used and also it fosters interoperability. So if you actually look at Kubernetes, for example, is a, is a wonderful example of how you actually find alignment through open source.

The second one is a, what we call open ecosystems that we really ought to foster common API is to bring more participants into the economy, which I think that incentivizes competition and actually forces the creation of API APIs that are sustainable and that have a longer timeless. And the last, but not least is designing architectures that promote agility and flexibility. And that is all about portability. It’s about interoperability, it is about this [00:30:00] centralized architecture is about the distributed architectures and it is about a collaboration of services. 

Alistair Croll: Awesome. Well, it sounds like you’re in the thick of all of these changes, which I think when we talk about like pod competing, we don’t realize how important that is. We kind of take it for granted. It’s just in the cloud and this was the original definition of ubiquitous computing was competing that just kind of received. Right. It became invisible and that’s true of internet things, but also of cloud computing. But as it becomes invisible, I think it’s really important to take take stock of it’s inherent bias is its inherent footprint costs, its inherent lock-ins and the other concerns so that we are moving towards ubiquitous computing that benefits everybody. So really fascinating to hear this and I think. Last word for you: when it comes to trade-offs and this is, you’re going to have some time to think about this, I guess when it comes to trade-offs what is the biggest trade-off we’re not talking about? And I’m gonna give you a pause to think about that because we have time to do and then we [00:31:00] can edit this long pause out, but you talked about three trade-offs, right. You talked about sovereignty versus sameness, I guess, you talked about ubiquity versus the footprint and cost of that and then you talked about privacy versus effectiveness, and I tried to summarize this. I hope we got it right. Is there a fourth one? Like what’s the other trade-off we’re not talking about that we should have. 

Patricia Florissi: I would say that I would not pick any one of the three that we talked. I think that the trade off that we are not talking about is how can we continue the growth of cloud in a sustainable way, as you said, it becomes so transparent that is part of the fabric, right? This is where through innovation requires where while at the same time, We are imposing on enterprises, on developers so many areas of concern, and I’ll give you four. Privacy,security. It’s about sustainability and it is about sovereignty. So [00:32:00] it’s not only about how can I develop innovative technologies anymore, but it’s how do I become how do I attend all of those four constraints? How do I do that in the presence of those four constraints while at the same time requiring the world accelerates the speed of innovation. 

Alistair Croll: Yeah, so the biggest trade-off is managing the trade-offs properly versus growing fast, which is very meta and that sounds like a great way to end the conversation. So Patricia, thank you much. It was an absolute delight talking about this with you. It makes me feel great. I, I, first of all, didn’t know that my searches have no carbon footprint, so I’m going to go search some stuff now. But thank you so much. This was a real pleasure. And again, thank you to Google for supporting FWD50 and helping us to put on the event overall and for these trade offs, which has given us lots to think about. 

Patricia Florissi: Our pleasure. Thank you.[00:33:00]

Technology is a series of tradeoffs. Blockchain gives us permanence in a digital world—at a large energy cost. Cloud computing gives us on-demand capacity—as long as we share it with others. And algorithms can help us navigate the world—if we provide access to information to learn from.

Patricia Florissi is a technical director in the office of the CTO at Google Cloud. She wrestles with these tradeoffs on a regular basis, something I had a chance to learn about in a recent conversation for the FWD50 Industry Innovations series.

Patricia outlined three distinct tradeoffs as computing becomes ubiquitous, and how Google Cloud is addressing them:

  • We want data privacy but also AI to be as free of unfair bias as possible. On one hand, we want technology that works for everyone – aligned with human rights, in areas such as privacy, unintended bias, and inclusion. But on the other hand, it is necessary to have access to appropriate data in order to test, validate, and train models.
  • The carbon footprint needed to make computing ubiquitous. Google uses offsets to make Google Cloud—and all Google services—carbon neutral.
  • The loss of data sovereignty and autonomy when putting all the compute workloads into a single cloud. A focus on open source, data liberation, and workload portability helps mitigate the risks here.

While most of the conversation was technical, it delved into philosophy somewhat. These tensions can be between the individual (customization, convenience, and sovereignty) and the collective (maximized utility for the group, climate change, and shared efficiency.) We also touched on the role of governments in regulating these tradeoffs with policies and incentives, and in adopting new platforms for digital transformation.