ONE DAY IN MARCH 2023, Arati Prabhakar brought a laptop into the Oval Office and showed the future to Joe Biden. Six months later, the president issued a sweeping executive order that set a regulatory course for AI.
This all happened because ChatGPT had stunned the world. In an instant it became very, very obvious that the United States needed to speed up its efforts to regulate the AI industry-and adopt policies to take advantage of it.
While the potential benefits were unlimited (Social Security customer service that works!), so were the potential downsides, like floods of disinformation or even, in the view of some, human extinction. Someone had to demonstrate that to the president.
The job fell to Prabhakar, because she is the director of the White House Office of Science and Technology Policy and holds cabinet status as the president's chief science and technology adviser; she'd already been methodically educating top officials about the transformative power of AI. But she also has the experience and bureaucratic savvy to make an impact with the most powerful person in the world.
Born in India and raised in Texas, Prabhakar has a PhD in applied physics from Caltech and previously ran two US agencies: the National Institute of Standards and Technology and the Department of Defense's Advanced Research Projects Agency. She also spent 15 years in Silicon Valley as a venture capitalist, including as president of Interval Research, Paul Allen's legendary tech incubator, and has served as vice president or chief technology officer at several companies.
Prabhakar assumed her current job in October 2022-just in time to have AI dominate the agenda-and helped to push out that 20,000-word executive order, which mandates safety standards, boosts innovation, promotes AI in government and education, and even tries to mitigate job losses. She replaced biologist Eric Lander, who had resigned after an investigation concluded that he ran a toxic workplace. Prabhakar is the first person of color and first woman to be appointed director of the office.
We spoke at the kitchen table of Prabhakar's Silicon Valley condo-a simply decorated space that, if my recollection is correct, is very unlike the OSTP offices in the ghostly, intimidating Eisenhower Executive Office Building in DC. Happily, the California vibes prevailed, and our conversation felt very unintimidating even at ease. We talked about how Bruce Springsteen figured into Biden's first ChatGPT demo, her hopes for a semiconductor renaissance in the US, and why Biden's war on cancer is different from every other president's war on cancer.
I also asked her about the status of the unfilled role of chief technology officer for the nation-a single person, ideally kind of geeky, whose entire job revolves around the technology issues driving the 21st century.
STEVEN LEVY: Why did you sign up for this job? ARATI PRABHAKAR: Because President Biden asked. He sees science and technology as enabling us to do big things, which is exactly how I think about their purpose.
What kinds of big things?
The mission of OSTP is to advance the entire science and technology ecosystem. We have a system that follows a set of priorities. We spend an enormous amount on R&D in health. But both public and corporate funding are largely focused on pharmaceuticals and medical devices, and very little on prevention or clinical care practices-the things that could change health as opposed to dealing with disease. We also have to meet the climate crisis. For technologies like clean energy, we don't do a great job of getting things out of research and turning them into impact for Americans. It's the unfinished business of this country.
It's almost predestined that you'd be in this job. As soon as you got your physics degree at Caltech, you went to DC and got enmeshed in policy.
Yeah, I left the track I was supposed to be on. My family came here from India when I was 3, and I was raised in a household where my mom started sentences with, "When you get your PhD and become an academic..." It wasn't a joke. Caltech, especially when I finished my degree in 1984, was extremely ivory tower, a place of worship for science.
I learned a tremendous amount, but I also learned that my joy did not come from being in a lab at 2 in the morning and having that eureka moment.
Just on a lark, I came to Washington for, quote-unquote, one year on a congressional fellowship. The big change was in 1986, when I went to Darpa as a young program manager. The mission of the organization was to use science and technology to change the arc of the future. I had found my home.
How did you wind up at Darpa?
I had written a study on microelectronics R&D. We were just starting to figure out that the semiconductor industry wasn't always going to be dominated by the US. We worked on a bunch of stuff that didn't pan out but also laid the groundwork for things that did. I was there for seven years, left for 19, and came back as director. Two decades later the portfolio was quite different, as it should be. I got to christen the first self-driving ship that could leave a port and navigate across open oceans without a single sailor on board. The other classic Darpa thing is to figure out what might be the foundation for new capabilities. I ended up starting a Biological Technologies Office. One of the many things that came out of that was the rapid development and distribution of mRNA vaccines, which never would have happened without the Darpa investment.
One difference today is that tech giants are doing a lot of their own R&D, though not necessarily for the big leaps Darpa was built for.
Every developed economy has this pattern. First there's public investment in R&D. That's part of how you germinate new industries and boost your economy.
As those industries grow, so does their investment in R&D, and that ends up being dominant. There was a time when it was sort of 50-50 public-private. Now it's much more private investment. For Darpa, of course, the mission is breakthrough technologies and capabilities for national security.
Are you worried about that shift?
It's not a competition! Absolutely there's been a huge shift. That private tech companies are building the leading edge LLMS today has huge implications. It's a tremendous American advantage, but it has implications for how the technology is developed and used. We have to make sure we get what we need for public purposes.
Is the US government investing enough to make that happen?
I don't think we are. We need to increase the funding. One component of the AI executive order is a National AI Research Resource. Researchers don't have the access to data and computation that companies have. An initiative that Congress is considering, that the administration is very supportive of, would place something like $3 billion of resources with the National Science Foundation.
That's a tiny percentage of the funds going into a company like OpenAl.
It costs a lot to build these leading-edge models. The question is, how do we have governance of advanced AI and how do we make sure we can use it for public purposes? The government has got to do more. We need help from Congress.
But we also have to chart a different kind of relationship with industry than we've had in the past.
What might that look like?
Look at semiconductor manufacturing and the CHIPS Act.
We'll get to that later. First let's talk about the president. How deep is his understanding of things like Al?
Some of the most fun I've gotten on the job was working with the president and helping him understand where the technology is, like when we got to do the chatbot demonstrations for the president in the Oval Office.
What was that like?
Using a laptop with ChatGPT, we picked a topic that was of particular interest.
The president had just been at a ceremony where he gave Bruce Springsteen the National Medal of Arts. He had joked about how Springsteen was from New Jersey, just across the river from his state, Delaware, and then he made reference to a lawsuit between those two states. I had never heard of it. We thought it would be fun to make use of this legal case. For the first prompt, we asked ChatGPT to explain the case to a first grader. Immediately these words start coming out like, "OK, kiddo, let me tell you, if you had a fight with someone..." Then we asked the bot to write a legal brief for a Supreme Court case. And out comes this very formal legal analysis. Then we wrote a song in the style of Bruce Springsteen about the case.
We also did image demonstrations. We generated one of his dog Commander sitting behind the Resolute desk in the Oval Office.
So what was the president's reaction?
He was like, "Wow, I can't believe it. could do that." It wasn't the first time he was aware of AI, but it gave him direct experience. It allowed us to dive into what was really going on. It seems like a crazy magical thing, but you need to get under the hood and understand that these models are computer systems that people train on data and then use to make startlingly good statistical predictions.
There are a ton of issues covered in the executive order. Which are the ones that you sense engaged the president most after he saw the demo?
The main thing that changed in that period was his sense of urgency. The task that he put out for all of us was to manage the risks so that we can see the benefits. We deliberately took the approach of dealing with a broad set of categories. That's why you saw an extremely broad, bulky, large executive order. The risks to the integrity of information from deception and fraud, risks to safety and security, risks to civil rights and civil liberties, discrimination and privacy issues, and then risks to workers and the economy and IPthey're all going to manifest in different ways for different people over different timelines. Sometimes we have laws that already address those risks-turns out it's illegal to commit fraud! But other things, like the IP questions, don't have clean answers.
There are a lot of provisions in the order that must meet set deadlines. How are you doing on those?
They are being met. We just rolled out all the 90-day milestones that were met.
One part of the order I'm really getting a kick out of is the AI Council, which includes cabinet secretaries and heads of various regulatory agencies. When they come together, it's not like most senior meetings where all the work has been done. These are meetings with rich discussion, where people engage with enthusiasm, because they know that we've got to get AI right.
There's a fear that the technology will be concentrated among a few big companies. Microsoft essentially subsumed one leading startup, Inflection. Are you concerned about this centralization?
Competition is absolutely part of this discussion. The executive order talks specifically about that. One of the many dimensions of this issue is the extent to which power will reside only with those who are able to build these massive models.
The order calls for AI technology to embody equity and not include biases. A lot of people in DC are devoted to fighting diversity mandates. Others are uncomfortable with the government determining what constitutes bias. How does the government legally and morally put its finger on the scale?
Here's what we're doing. The president signed the executive order at the end of October. A couple of days later, the Office of Management and Budget came out with a memo-a draft of guidance about how all of government will use AI. Now we're in the deep, wonky part, but this is where the rubber meets the road. It's that guidance that will build in processes to make sure that when the government uses AI tools it's not embedding bias.
That's the strategy? You won't mandate rule...
This all happened because ChatGPT had stunned the world. In an instant it became very, very obvious that the United States needed to speed up its efforts to regulate the AI industry-and adopt policies to take advantage of it.
While the potential benefits were unlimited (Social Security customer service that works!), so were the potential downsides, like floods of disinformation or even, in the view of some, human extinction. Someone had to demonstrate that to the president.
The job fell to Prabhakar, because she is the director of the White House Office of Science and Technology Policy and holds cabinet status as the president's chief science and technology adviser; she'd already been methodically educating top officials about the transformative power of AI. But she also has the experience and bureaucratic savvy to make an impact with the most powerful person in the world.
Born in India and raised in Texas, Prabhakar has a PhD in applied physics from Caltech and previously ran two US agencies: the National Institute of Standards and Technology and the Department of Defense's Advanced Research Projects Agency. She also spent 15 years in Silicon Valley as a venture capitalist, including as president of Interval Research, Paul Allen's legendary tech incubator, and has served as vice president or chief technology officer at several companies.
Prabhakar assumed her current job in October 2022-just in time to have AI dominate the agenda-and helped to push out that 20,000-word executive order, which mandates safety standards, boosts innovation, promotes AI in government and education, and even tries to mitigate job losses. She replaced biologist Eric Lander, who had resigned after an investigation concluded that he ran a toxic workplace. Prabhakar is the first person of color and first woman to be appointed director of the office.
We spoke at the kitchen table of Prabhakar's Silicon Valley condo-a simply decorated space that, if my recollection is correct, is very unlike the OSTP offices in the ghostly, intimidating Eisenhower Executive Office Building in DC. Happily, the California vibes prevailed, and our conversation felt very unintimidating even at ease. We talked about how Bruce Springsteen figured into Biden's first ChatGPT demo, her hopes for a semiconductor renaissance in the US, and why Biden's war on cancer is different from every other president's war on cancer.
I also asked her about the status of the unfilled role of chief technology officer for the nation-a single person, ideally kind of geeky, whose entire job revolves around the technology issues driving the 21st century.
STEVEN LEVY: Why did you sign up for this job? ARATI PRABHAKAR: Because President Biden asked. He sees science and technology as enabling us to do big things, which is exactly how I think about their purpose.
What kinds of big things?
The mission of OSTP is to advance the entire science and technology ecosystem. We have a system that follows a set of priorities. We spend an enormous amount on R&D in health. But both public and corporate funding are largely focused on pharmaceuticals and medical devices, and very little on prevention or clinical care practices-the things that could change health as opposed to dealing with disease. We also have to meet the climate crisis. For technologies like clean energy, we don't do a great job of getting things out of research and turning them into impact for Americans. It's the unfinished business of this country.
It's almost predestined that you'd be in this job. As soon as you got your physics degree at Caltech, you went to DC and got enmeshed in policy.
Yeah, I left the track I was supposed to be on. My family came here from India when I was 3, and I was raised in a household where my mom started sentences with, "When you get your PhD and become an academic..." It wasn't a joke. Caltech, especially when I finished my degree in 1984, was extremely ivory tower, a place of worship for science.
I learned a tremendous amount, but I also learned that my joy did not come from being in a lab at 2 in the morning and having that eureka moment.
Just on a lark, I came to Washington for, quote-unquote, one year on a congressional fellowship. The big change was in 1986, when I went to Darpa as a young program manager. The mission of the organization was to use science and technology to change the arc of the future. I had found my home.
How did you wind up at Darpa?
I had written a study on microelectronics R&D. We were just starting to figure out that the semiconductor industry wasn't always going to be dominated by the US. We worked on a bunch of stuff that didn't pan out but also laid the groundwork for things that did. I was there for seven years, left for 19, and came back as director. Two decades later the portfolio was quite different, as it should be. I got to christen the first self-driving ship that could leave a port and navigate across open oceans without a single sailor on board. The other classic Darpa thing is to figure out what might be the foundation for new capabilities. I ended up starting a Biological Technologies Office. One of the many things that came out of that was the rapid development and distribution of mRNA vaccines, which never would have happened without the Darpa investment.
One difference today is that tech giants are doing a lot of their own R&D, though not necessarily for the big leaps Darpa was built for.
Every developed economy has this pattern. First there's public investment in R&D. That's part of how you germinate new industries and boost your economy.
As those industries grow, so does their investment in R&D, and that ends up being dominant. There was a time when it was sort of 50-50 public-private. Now it's much more private investment. For Darpa, of course, the mission is breakthrough technologies and capabilities for national security.
Are you worried about that shift?
It's not a competition! Absolutely there's been a huge shift. That private tech companies are building the leading edge LLMS today has huge implications. It's a tremendous American advantage, but it has implications for how the technology is developed and used. We have to make sure we get what we need for public purposes.
Is the US government investing enough to make that happen?
I don't think we are. We need to increase the funding. One component of the AI executive order is a National AI Research Resource. Researchers don't have the access to data and computation that companies have. An initiative that Congress is considering, that the administration is very supportive of, would place something like $3 billion of resources with the National Science Foundation.
That's a tiny percentage of the funds going into a company like OpenAl.
It costs a lot to build these leading-edge models. The question is, how do we have governance of advanced AI and how do we make sure we can use it for public purposes? The government has got to do more. We need help from Congress.
But we also have to chart a different kind of relationship with industry than we've had in the past.
What might that look like?
Look at semiconductor manufacturing and the CHIPS Act.
We'll get to that later. First let's talk about the president. How deep is his understanding of things like Al?
Some of the most fun I've gotten on the job was working with the president and helping him understand where the technology is, like when we got to do the chatbot demonstrations for the president in the Oval Office.
What was that like?
Using a laptop with ChatGPT, we picked a topic that was of particular interest.
The president had just been at a ceremony where he gave Bruce Springsteen the National Medal of Arts. He had joked about how Springsteen was from New Jersey, just across the river from his state, Delaware, and then he made reference to a lawsuit between those two states. I had never heard of it. We thought it would be fun to make use of this legal case. For the first prompt, we asked ChatGPT to explain the case to a first grader. Immediately these words start coming out like, "OK, kiddo, let me tell you, if you had a fight with someone..." Then we asked the bot to write a legal brief for a Supreme Court case. And out comes this very formal legal analysis. Then we wrote a song in the style of Bruce Springsteen about the case.
We also did image demonstrations. We generated one of his dog Commander sitting behind the Resolute desk in the Oval Office.
So what was the president's reaction?
He was like, "Wow, I can't believe it. could do that." It wasn't the first time he was aware of AI, but it gave him direct experience. It allowed us to dive into what was really going on. It seems like a crazy magical thing, but you need to get under the hood and understand that these models are computer systems that people train on data and then use to make startlingly good statistical predictions.
There are a ton of issues covered in the executive order. Which are the ones that you sense engaged the president most after he saw the demo?
The main thing that changed in that period was his sense of urgency. The task that he put out for all of us was to manage the risks so that we can see the benefits. We deliberately took the approach of dealing with a broad set of categories. That's why you saw an extremely broad, bulky, large executive order. The risks to the integrity of information from deception and fraud, risks to safety and security, risks to civil rights and civil liberties, discrimination and privacy issues, and then risks to workers and the economy and IPthey're all going to manifest in different ways for different people over different timelines. Sometimes we have laws that already address those risks-turns out it's illegal to commit fraud! But other things, like the IP questions, don't have clean answers.
There are a lot of provisions in the order that must meet set deadlines. How are you doing on those?
They are being met. We just rolled out all the 90-day milestones that were met.
One part of the order I'm really getting a kick out of is the AI Council, which includes cabinet secretaries and heads of various regulatory agencies. When they come together, it's not like most senior meetings where all the work has been done. These are meetings with rich discussion, where people engage with enthusiasm, because they know that we've got to get AI right.
There's a fear that the technology will be concentrated among a few big companies. Microsoft essentially subsumed one leading startup, Inflection. Are you concerned about this centralization?
Competition is absolutely part of this discussion. The executive order talks specifically about that. One of the many dimensions of this issue is the extent to which power will reside only with those who are able to build these massive models.
The order calls for AI technology to embody equity and not include biases. A lot of people in DC are devoted to fighting diversity mandates. Others are uncomfortable with the government determining what constitutes bias. How does the government legally and morally put its finger on the scale?
Here's what we're doing. The president signed the executive order at the end of October. A couple of days later, the Office of Management and Budget came out with a memo-a draft of guidance about how all of government will use AI. Now we're in the deep, wonky part, but this is where the rubber meets the road. It's that guidance that will build in processes to make sure that when the government uses AI tools it's not embedding bias.
That's the strategy? You won't mandate rule...
You're reading a preview of
WIRED (Digital) - 1 Issue, July - August 2024