00:00The ranking member of the full committee, Mr. Raskin, for his questions.
00:04Thanks, Mr. Chairman. Continuing with you, Dr. Villasenor,
00:07many of the AI models incorporate open source code, coding that's publicly available into
00:15their models. How does this affect their ability to assert a trade secret claim?
00:22Well, it's a really good question. I think it depends in part on how they've incorporated
00:27these open source sources, for example. But it is certainly possible to have a system that had
00:33some component that was based on open source, but that also has a lot of intellectual property
00:38that's created independent of that. And there would be trade secret value in that. And for
00:42example, on how they integrate that with open source information. So I wouldn't conclude or I
00:46wouldn't believe that the presence or the use of open source would eviscerate trade secret rights
00:51for the companies that use it. All right. Dr. Jensen, do you agree with that,
00:55that you can still maintain some kind of trade secret, some actionable trade secret claim,
01:02even if you've been using open source code as part of your programming?
01:06It's hard. And this is why, again, I want to stress that idea of going on the offense.
01:11And this is where it definitely gets outside the jurisdiction of this committee and gets really
01:14more into what's the rights of the executive. I think you are going to have to find ways to
01:19potentially sabotage the most malicious offenders that you identify as linked to a foreign government
01:25going after you. Because you can protect it. He's right with the larger trade secret. But then it
01:29becomes only a matter of time before they put the pieces together. And if you have someone engaged in
01:34industrial size, like generation defining theft, you can't trust that person. So where is the balance
01:41there of how you actually take other measures that are getting outside this jurisdiction to start to
01:45hold those countries accountable and then undermine it? That's the only way I think it's going to work.
01:50Okay. Ms. Turner, you've advocated for transparency requirements for developers of
01:55high-stakes AI systems, including transparency for training data, capability testing, safety testing,
02:03risk management, safety cases, and so on. What does transparency mean in this context?
02:09And what is its purpose in these different areas?
02:12It's a great question. I would just clarify, I would not necessarily advocate for transparency over
02:16training data, but the rest of the items that you listed, certainly. I think there's different
02:20purposes and different forms it can take. So the simplest form of transparency is transparency to
02:24the public. I think this has a very broad range of benefits. It means that it's already transparent
02:28to governments, because governments can access anything that is public. But it also lets a broader
02:32set of academics, experts, employees of companies who might be affected. It just really empowers the
02:39public to understand the stakes of this technology and how the technology is changing over time,
02:43which I think is important, given that it is going to affect all of our lives very profoundly and
02:47perhaps in sudden and unexpected ways. A different form of transparency is transparency
02:52to government. So there's some kinds of information that maybe are not suited to be
02:56public. Maybe we want to keep them protected. And that could particularly be relevant for
03:01national security-relevant information. So, for example, I think a big area of concern right now is
03:07the potential cybersecurity capabilities of the most advanced models, whether they're going to
03:12make critical infrastructure systems more vulnerable than they have been in the past.
03:15Maybe that's not something that we want disclosed publicly immediately, but it is an area where
03:19partnership with government makes sense. A third form of transparency, which is maybe
03:23transparency is not quite the right word, but I think it's important to also think about
03:26the ability to bring in independent third parties into these companies and provide independent
03:31oversight or independent verification of some high-stakes claims they're making. So if they're saying
03:35that they're running certain kinds of tests, that they're performing certain kinds of safety
03:38protocols, can you bring in a third party under NDA to work inside the company's own
03:42facilities? That might allow you to get a little bit of additional independent verification so you're
03:47not just taking the company's word for everything for these high-stakes claims they're making.
03:50Great. Thank you. Dr. Jensen, the Washington Post reported that tech companies or some tech companies
03:58are telling their employees on visas not to leave the United States out of concern that they won't be
04:03readmitted to the country or they're going to be stopped at the border. Visa holders are obviously
04:08running a lot of tech companies. They're working at startups. They're key in the field. Seventy percent
04:16of graduate students in AI-related fields are believed to be international students. So if the
04:21Trump administration begins to deny H-1B visas, as he did in the first term, what will the consequences
04:30be on American AI innovation and on these businesses? Well, I'll start with this combat veteran. It'll cut my lab
04:35in half. Literally, of the three people I employ, I would cut it at four people counting me total. I
04:42would lose two of them. And I think I'm a microcosm compared to the size and scale of how this could
04:48ripple through the American ecosystem. That doesn't mean that we don't have the right and the executive
04:52doesn't have the right to determine how best we seal our borders and determine very complex
04:57immigration questions. But I think we have to have a really hard conversation as a nation about
05:02what really is the major competition of the 21st century. If it truly is AI, and AI becomes the
05:07focal point of grand strategy, then these immigration discussions have to be taken as part of that.
05:13And we're going to have to find some areas of bipartisan agreement or compromise. Or we're going to have
05:17to really double down an investment in STEM education, which is going to even take longer,
05:22and we have to figure out how it works. Now, I just say one note quick about transparency to double
05:26down. There's a really interesting point of transparency. So my lab, we do benchmarking.
05:30And you're going to need almost a Tocquevillian civil society function, where whether it's just
05:36naturally happens or it's incentivized by the people's house, independent groups, not government,
05:42not business, come together and hold company models accountable. So in our case, we actually do the hard
05:48work to say when a DeepSeq or a Gemini responds with escalatory responses on a data set trained on
05:56standard international relations and foreign policy decision making. That's just one example.
06:00So it's not just transparency in the ways we've discussed. It's also let's get creative and think
06:05about all of American society and what are the ways civil society can be incentivized to organize
06:11organically from below and provide checks on us. Because I don't think benchmarking can just be the
06:16government. That'll lead to overreach. And it can't just be companies because that'll lead to kind of
06:20skirting claims. And there's perverse incentives for either the government or just firms. So transparency
06:25needs to also include civil society. And to us thinking about what does that benchmarking look
06:30like conducted through voluntary American associations?
06:33I appreciate that. I yield back, Mr. Chairman.
