Day 1 Opening Remarks
Summary
Adam Gleave challenges the perceived dichotomy between AI innovation and safety, presenting technical solutions and stressing the need for cross-sector collaboration.
SESSION Transcript
Hello everyone. It is my pleasure to welcome you all to the inaugural Technical Innovations for AI Policy Conference. I'm Adam Gleave. I'm the co-founder and chief executive of FAR.AI, and I'll be delivering some brief opening remarks before handing the stage over to our many wonderful speakers.
So I often see debates around AI policy framed as this dichotomy between innovation—rapid technological progress on the one extreme—and safety—onerous, really expensive regulation on the other hand. And I can see why this is how it's framed, because right now I would say we're simply lacking effective technical mechanisms needed to be able to develop good AI policy. But I believe that through technical innovation, we can overcome this binary choice between two undesirable outcomes and enable both progress and trustworthy development. And that's the goal of this event: to catalyze development of better tools needed to implement these effective policies.
Now this might sound daunting, but the good news is that we have solved similar problems before with many novel technologies. So for well over 100 years, air pollution caused problems like acid rain. And this was a tricky problem to solve because using some blunt policy instrument like "ban all coal plants" would cause significant economic damage. But just doing nothing also didn't seem like an effective option from a public health perspective. But once standardized measurement techniques were developed, this changed everything. So today, hundreds of power plants in the United States have continuous emissions monitoring systems installed that in real time measure how many emissions that power plant is making. And then these power plants have a cap that they can trade with other power plants. And this enables utility operators to choose the right approach for them. Do they idle the power plant except at peak demands? Do they retrofit filters? Do they buy extra emissions permits from other utilities? So this solves the public health problem while minimizing cost to industry and actively incentivizing innovation.
A story a little bit closer to home for AI policy would be differential privacy. This is a mathematical framework that lets you anonymize data, protecting individual privacy while still being able to calculate in aggregate relevant statistics. This is used by the US Census Bureau; it's used by companies like Apple and Google. And it solved what previously seemed like an intractable policy problem: How do you balance the utility of making available data for public analysis with protecting individuals' privacy?
Now it would be overly optimistic to say that every problem has such a clean technical solution, but I believe there are many areas where technical innovations can directly help policy, or at least help inform it. One example I'd like to highlight is that right now we are simply lacking mechanisms to securely evaluate models by third parties. And a particularly important area where we need this would be in chemical, biological, radiological, and nuclear weapon risk evaluation, or CBRN risks. And a lot of expertise here is locked up in national labs and it's classified. And so when I've spoken to people in these areas, we often simply cannot test frontier models because that would be sending these data over public, unclassified Internet. And this is a very solvable problem. You simply put the models in a secure testing environment run by some trusted third party. There's already a number of cloud providers that have the relevant certifications. And this is a way in which you can maintain the security of that benchmark data while also protecting the commercially sensitive model weights' intellectual property from being leaked.
Another example that's gaining increasing momentum would be on-chip mechanisms to let you monitor and in some cases restrict usage of particular AI chips. And, you know, one immediate problem this could help with is export control. So today the most advanced AI accelerator chips are subject to export controls. But a recent white paper coming out from the Center for New American Security estimates that over 100,000 of Nvidia's most advanced chips have been smuggled in the last 12 months into China in contravention of these export controls. And that's worth around $2.5 billion. So it's quite a sizable business. And it's in some ways not surprising because these export controls run on paperwork and trust. So it's very easy to export it to a third country, have a shell company export it from there into China. But you can have on-chip mechanisms that would let you detect the location of a chip and potentially even geofence its deployments—will stop functioning outside of a certain area.
This isn't that different from what people are already doing today with various hardware chips for commercial reasons. So there's been a lot of focus on supply chain security. There are root of trust mechanisms like OpenTitan that validate firmware that a system is loading. The boot hasn't been tampered with at any point in the supply chain. And from then on you can have this completely secure boot all the way up to the operating system and the application. So you can build on these mechanisms to have actually technically verifiable governance mechanisms.
So why we've brought this assorted group together is that we currently have to choose between these unsatisfactory policy choices that's resulting in endless circular debates between innovation and safety, and technical innovations can open up completely new interventions. But to achieve this, we're going to need input from a variety of policymakers to make sure that technical researchers are actually working on real-world problems. So it's no coincidence that we chose DC to have this event, but we also need a variety of technical experts from academia, nonprofits, industry to develop and deploy it. So you can usually pretty easily tell based on someone wearing a suit or a T-shirt which side of the aisle someone's on, so you won't have any trouble finding each other. But I would encourage you to talk to new people with a different background so that we do have that kind of cross-fertilization.
So I often see debates around AI policy framed as this dichotomy between innovation—rapid technological progress on the one extreme—and safety—onerous, really expensive regulation on the other hand. And I can see why this is how it's framed, because right now I would say we're simply lacking effective technical mechanisms needed to be able to develop good AI policy. But I believe that through technical innovation, we can overcome this binary choice between two undesirable outcomes and enable both progress and trustworthy development. And that's the goal of this event: to catalyze development of better tools needed to implement these effective policies.
Now this might sound daunting, but the good news is that we have solved similar problems before with many novel technologies. So for well over 100 years, air pollution caused problems like acid rain. And this was a tricky problem to solve because using some blunt policy instrument like "ban all coal plants" would cause significant economic damage. But just doing nothing also didn't seem like an effective option from a public health perspective. But once standardized measurement techniques were developed, this changed everything. So today, hundreds of power plants in the United States have continuous emissions monitoring systems installed that in real time measure how many emissions that power plant is making. And then these power plants have a cap that they can trade with other power plants. And this enables utility operators to choose the right approach for them. Do they idle the power plant except at peak demands? Do they retrofit filters? Do they buy extra emissions permits from other utilities? So this solves the public health problem while minimizing cost to industry and actively incentivizing innovation.
A story a little bit closer to home for AI policy would be differential privacy. This is a mathematical framework that lets you anonymize data, protecting individual privacy while still being able to calculate in aggregate relevant statistics. This is used by the US Census Bureau; it's used by companies like Apple and Google. And it solved what previously seemed like an intractable policy problem: How do you balance the utility of making available data for public analysis with protecting individuals' privacy?
Now it would be overly optimistic to say that every problem has such a clean technical solution, but I believe there are many areas where technical innovations can directly help policy, or at least help inform it. One example I'd like to highlight is that right now we are simply lacking mechanisms to securely evaluate models by third parties. And a particularly important area where we need this would be in chemical, biological, radiological, and nuclear weapon risk evaluation, or CBRN risks. And a lot of expertise here is locked up in national labs and it's classified. And so when I've spoken to people in these areas, we often simply cannot test frontier models because that would be sending these data over public, unclassified Internet. And this is a very solvable problem. You simply put the models in a secure testing environment run by some trusted third party. There's already a number of cloud providers that have the relevant certifications. And this is a way in which you can maintain the security of that benchmark data while also protecting the commercially sensitive model weights' intellectual property from being leaked.
Another example that's gaining increasing momentum would be on-chip mechanisms to let you monitor and in some cases restrict usage of particular AI chips. And, you know, one immediate problem this could help with is export control. So today the most advanced AI accelerator chips are subject to export controls. But a recent white paper coming out from the Center for New American Security estimates that over 100,000 of Nvidia's most advanced chips have been smuggled in the last 12 months into China in contravention of these export controls. And that's worth around $2.5 billion. So it's quite a sizable business. And it's in some ways not surprising because these export controls run on paperwork and trust. So it's very easy to export it to a third country, have a shell company export it from there into China. But you can have on-chip mechanisms that would let you detect the location of a chip and potentially even geofence its deployments—will stop functioning outside of a certain area.
This isn't that different from what people are already doing today with various hardware chips for commercial reasons. So there's been a lot of focus on supply chain security. There are root of trust mechanisms like OpenTitan that validate firmware that a system is loading. The boot hasn't been tampered with at any point in the supply chain. And from then on you can have this completely secure boot all the way up to the operating system and the application. So you can build on these mechanisms to have actually technically verifiable governance mechanisms.
So why we've brought this assorted group together is that we currently have to choose between these unsatisfactory policy choices that's resulting in endless circular debates between innovation and safety, and technical innovations can open up completely new interventions. But to achieve this, we're going to need input from a variety of policymakers to make sure that technical researchers are actually working on real-world problems. So it's no coincidence that we chose DC to have this event, but we also need a variety of technical experts from academia, nonprofits, industry to develop and deploy it. So you can usually pretty easily tell based on someone wearing a suit or a T-shirt which side of the aisle someone's on, so you won't have any trouble finding each other. But I would encourage you to talk to new people with a different background so that we do have that kind of cross-fertilization.
