Full interview: Palantir CTO Shyam Sankar

[0:00] And we're joined now by the chief technology officer of Palantir, Shyam Sankar. Welcome to [0:05] Face the Nation. Thank you for having me, Margaret. It's good to have you here in person. [0:08] We've been talking a lot about technology, artificial intelligence, and public perception [0:13] of it right now. And that's where I want to start with you, because our polling shows there's a [0:17] problem here. 50% of those polled by CBS say AI will create more problems than it solves, [0:23] 29% say it'll solve more than it creates, and 21% just really don't know. Like 45% of [0:29] Americans want the government to restrict use and development of AI. Is your position [0:35] there should not be any guardrails at this moment? No, not at all. That's not my position. I think [0:42] the burden is absolutely on the technology industry to show the American people that there are real [0:47] benefits to AI and not kind of ephemeral sort of benefits like it can help me schedule travel, [0:53] but actually meaningful benefits that result in job growth and prosperity for the American people. [0:57] I think one of the things that's really complicated in the current discussion is that there's an [1:02] unseen force driving the narrative around AI safety. And it's really this millenarian doomsday [1:09] cult meets messianic complex called effective altruism. The most famous effective altruist that [1:15] most Americans know is Sam Bankman Freed. And the effective altruists are, it's almost like a religion [1:21] that they've been thinking about doomsday scenarios from AI since the early 2000s. [1:27] He was, he is jailed. He's was the crypto investor. [1:31] Correct. [1:31] Who defrauded investors. But this sounds very academic. Like what makes you so focused on this? [1:39] Because there's only really roughly 2000 AI researchers in America, and the vast majority of [1:44] them believe in this sort of worldview around effective altruism. And if you hear someone like [1:50] Jacob Coxon telling you that AI has a 10% chance of ending humanity, and you believe- [1:56] The researcher who had this warning recently. [1:57] Yes, correct. And you believe he's speaking as a technically informed individual who does this work, [2:03] it's a scientific critique. I think the American people would be right to be quite panicked and [2:08] worried about what's going on here. But if you understand that it's actually a theological critique, [2:12] then you can, you can kind of wait against the empirical facts of what is actually happening [2:17] out there. You know, I take, I think there's a profound truth to the Pascal saying that every [2:21] human has a God-shaped hole in their heart. And it's important to understand that many of these [2:25] folks in Silicon Valley who are building this technology have filled that hole with the machine [2:30] God of AGI. And I mean, there's so many interesting stories around this. Like for example, [2:35] so many AI researchers have started smoking as a way of signaling to other researchers how much they [2:41] believe in AI. What? That makes no sense. What do you think? [2:44] I believe so much in AI that it's going to cure cancer. So I will, I'm showing you my chips are [2:49] all in. So there is kind of this mass psychosis social delusion around it. And I think it's getting [2:54] in the way of actually what is happening with AI? Where is it meeting the American worker in industry? [2:59] What are the results of doing that? And the doomsdayers have basically been wrong everything [3:03] they predicted. They told us that chat GPT2 was too dangerous to release. They told us at this point in the [3:09] timeline, we'd already have mass unemployment and job displacement. And the economists see that [3:14] actually we're having job growth in areas that we wouldn't have predicted doing exactly that. [3:18] So the technology industry has two burdens. One is to show very clearly to the American people [3:24] that AI is going to result in economic prosperity and security for the nation. And two, to actually [3:31] do and absorb the burden of engineering. You cannot abdicate this idea that designing safe products [3:38] is your job. You can't hand wave around it. If you look at the kind of narratives that come out of [3:44] the doomsayers, they seem like two contradictory narratives. One narrative is this technology is [3:50] going to end humanity. And simultaneously they say, but I believe it has some chance of creating [3:55] utopia. So at first blush, this seems contradictory, but it has one thing in common, which is that you [4:00] as a human have no agency. The machine God is going to decide the future. And only perhaps a small number [4:06] of us can try to avoid the horrible outcome. And I think it obfuscates the more profound reality that [4:11] we have a huge amount of human agency over how we're going to use this technology to deliver the [4:16] outcomes the American people expect. Well, let's talk about that because the other part of this [4:19] argument would be, and no one who wants, the people who don't want guardrails are simply in it for [4:25] commercial reasons. They want to keep selling chips or they just want access to market. So that would be the [4:29] counter argument to the idea that, you know, what you're saying, which is an ideological opposition [4:36] to guardrails. Maybe we could just back up and who's on the side of no guardrails? [4:41] Well, when I want to get into the specifics here, because even just defining what guardrails is [4:46] seems really complicated. And there's not even common agreement on that right now. But Jay Clayton, [4:52] who is the new DNI and the new AIs are said this week, the focus is on control and monitoring of [4:57] technologies with the potential to do harm. So what does that look like to you? Because it sounds [5:02] like he's not even sure yet. Well, I think there's a very clear view that we should be regulating the [5:07] application of the technology. So if you're going to use this in medicines, you know, the FDA, [5:12] as a republic, we have 250 years of rich laws in history on how to manage liability. I don't want [5:18] to trivialize the metaphor, but, you know, if I buy a car and the car is defective, that's the car [5:24] maker's problem. If I then use that car to go drive into my neighbor's house, that's my problem. [5:30] And, you know, this sort of obfuscation, the narrative is this technology is a total [5:35] difference of kind. I don't think so. I think it's a difference of degree. [5:39] Well, in that humans aren't in control of it, it would be a different kind. [5:44] No, humans are in control of it. Even if we look at the kind of salacious reporting around hugging [5:48] face, the models did not do anything they were not instructed to do. You know, they were, [5:53] the real, first you have human error. They were not in a sandbox. They had access to the internet. [5:57] It was a horrible cybersecurity configuration. Then from that perspective, they were asked to [6:02] succeed at a specific benchmark, and they were not given constraints on how to do that. And they [6:08] said, oh, the best way to succeed at the benchmark is to get the answers to the test and to do so. [6:12] So for that, who should take action to? [6:15] I think it's a defective product. Like, if you were going to sell that. [6:18] For the Justice Department to then bring suit, like, on the programmers? Like, who's liable for [6:23] the technology not, you know, staying within the sandbox? [6:27] Of course. In that specific case, it would be the AI companies. If another company had licensed [6:32] that technology to use in another, in an application, and the product worked as expected, [6:36] but you misused the product, then it would be you that are responsible for that. [6:40] So you believe that, in my understanding of you, the existing laws on the books should be able to [6:47] handle technology as it develops? I think it is a massive running head start. I think the laws on [6:52] the books speak to a lot of the current considerations. Might we need a few more [6:55] things on the margin? Sure. But I think we'll figure that out as we go through it. [6:58] So let's talk about some of those things, because you were with the president when he summoned some [7:01] of these AI leaders recently, and he asked them all to sign a pledge to implement robust internal [7:07] controls, partner with an external auditor, and establish a committee to evaluate reports from [7:11] auditors. That was Google, Anthropic, OpenAI, NVIDIA, Elon Musk. They all signed it. Palantir, [7:16] your company didn't. Why? Well, those accords were for companies that actually make AI models, [7:21] which we do not. But you helped to use them and build out platforms. Like, you don't see [7:27] a reason for constraints there. Oh, we do have those constraints. In fact, [7:32] that part of that's our entire business is how do you safely deploy these models in specific use cases? [7:36] We're just not a party. We weren't asked to be a party of the accord. [7:38] But you were there at that meeting. I was. [7:40] So what happened? Like, what was the conversation? [7:43] Well, the conversation was really an assertion that the American people expect AI to generate [7:49] economic prosperity and to do so securely, and that we as companies have an affirmative [7:53] obligation to deliver on exactly those things, and that we have existing laws to look at this, [7:59] and that we should be all taking this quite seriously. It was a very productive conversation. [8:02] Because it sounded like the White House was saying something more should be done, [8:05] and now you have this new AI, Zarr J. Clayton, going out to Silicon Valley this week. He said [8:10] he's launched this 120-day review of AI policy. What do you hope comes from that? Or is this [8:16] another one of these, like, blue-ribbon commissions that doesn't produce anything? [8:19] No, I think a lot will come up to it, which is really focus and energy and using the existing [8:23] agencies and regulatory apparatus to make sure we're on top of this stuff and providing support [8:27] where it's needed. I think one of the challenges, going back to my comments on effective [8:31] altruism is today, the so-called independent research organizations that are overseeing [8:36] this are not really independent. They have conflicts with the labs. They are made up of [8:41] people with a singular worldview around effective altruism, looking for the doomsday scenarios. [8:46] You know, you have evaluators who are essentially pointing a loaded gun at a target, saying, [8:50] I can't believe the gun worked, when they should really be saying, what guardrails do we have in [8:54] place? Are these products working to the specifications designed, and what are the opportunities for [9:01] misuse? So I've spoken to people, for example, in the financial industry who say, we did need [9:06] some regulation here, but there's the SEC. There is, you know, the CFTC. They oversee certain [9:12] operations here. Don't you think within the current institutions, there could be some sort of [9:18] entity that oversees technology? I think there's going to need to be an incident response and [9:26] reporting regime that creates more visibility and transparency into what you're learning in the [9:30] testing well ahead of product launches that give visibility and transparency to the government [9:35] of how people are doing and implementing exactly the existing laws. Who does that? Or who could do [9:40] that? I will defer to Director Clayton on that, but I think that's one of the things that I suspect [9:45] they will be driving towards. He's got to figure that out. Well, we look at the [9:50] last administration, President Biden and Xi Jinping came to, like, a very limited agreement, [9:55] but at least on one principle, which was humans, not AI, should retain control over decisions to [10:01] use nuclear weapons. Would it be sensible? Well, that seems obvious. Well, that seems terrifying, [10:07] first of all, that they had to come to a conversation about that, but shouldn't that be [10:11] expanded elsewhere in the defense space? Would you agree with that? Well, the current department [10:15] policy is, there's two parts of it. There's the kind of military doctrine, which is that a human is [10:19] always accountable for every decision that's made. And the other part of it is that humans are in [10:24] the decision-making processes that relate to lethal force here. [10:29] But in terms of the things where you think there should be more guardrails, is the defense space [10:36] one of them? Biotech, I've heard others. All spaces need to have a thought around safety [10:42] of the application. Mm-hmm. But that's still, who does that and how? Well, I think the department [10:49] and the Congress should do it with defense. The Department of Transportation should do it [10:54] with transportation. The FDA should do it with medicines. Mm-hmm. So watch for that possibly [10:59] happen. Because, you know, I hear from people in the technology space this argument that we need to [11:07] just go so furiously forward to compete with China. China is pushing for AI to be integrated into all [11:14] of its industries, but they have the ability to also delay things, to get governance to catch up. [11:21] They've done that. They've tapped the brakes themselves before. So do you think the U.S. [11:27] should also consider tapping the brakes in certain areas, just slow things until they have [11:31] some idea of the guardrails? You know, we have this beautiful, messy process in America, but I would [11:37] argue, as frustrating as it might look, it's kind of working. So we have frontier models now that have [11:42] not been released because people have realized maybe we're not ready to release them. That would [11:46] have been an unimaginable thought a year ago. Mm-hmm. And so as people are starting to wrap their heads [11:51] around, whether it's the labs, government, I think it is a consequence of our messy process and the [11:57] discussions that people are having that we're evolving how we're doing this. And the pace is [12:01] naturally changing. Mm-hmm. Um, let's talk a little bit about DOD, the Department of War, [12:08] Department of Defense. Um, they have recently ceased using Anthropic products as part of this dispute [12:13] over the firm's insistence on safeguards for military use of AI. I know you're aware of this case. Um, [12:20] Anthropics, Claude, was part of the Maven platform, which Palantir uses. How much of an impact will this [12:27] decision here have and who replaces them? Well, the department today has a wide range of model [12:34] providers, everything from XAI and OpenAI to open weight models like NemoTron from NVIDIA. So I [12:42] don't think, um, I don't, there's been no operational impact from that. I think there's a kind of a [12:47] broader statement, which really has been powering the growth of our commercial business, which is [12:50] companies are realizing, and it equally applies to the government, that you're not going to pick any [12:55] one model provider, that actually your sovereignty and agency as an institution comes down to your [13:00] ability to kind of control, decide which models you're going to use in what context, have the [13:04] ability to switch models, to be able to control the weights themselves, where you need to feel [13:08] that control over them. It is your alpha. It is what makes your institution special. [13:13] So you don't know yet who's going to replace them, but you don't see a big impact from this kind [13:19] of decision. It was certainly a huge story for Silicon Valley and the defense space to hear that- [13:25] It's already been replaced. It's already, it's already been replaced and, and users who are [13:28] building AI applications get to pick the models that they're using. It's a suite of models that [13:32] are available from, you know, Google, from XAI, from OpenAI. [13:38] You have said recently that the war in Iran is the first major conflict where AI played a central [13:43] role, particularly for targeting. Do you see risks for that? [13:48] That's, that's, well, you know, targeting is this, it's kind of this very salacious word. Really, [13:52] targeting is a process. And if you think about the military operations, you know, the ability to [13:58] execute a thousand strikes in a single day would have taken two weeks of planning by 50 people [14:03] before. And being able to do that with one person in four hours provides you a significant battlefield [14:09] advantage against your adversaries. So it's really not about, you know, people jump immediately to [14:13] something like AI picking the targets. That's not what's happening. It's really given a set of effects [14:18] that you want to create. I want to be able to take out the enemy's air defenses. Okay. Um, what targets [14:22] relate to that? Okay. How do I then push them through a process where I understand are any of [14:27] these targets on the no strike list? Are any of these targets too close to civilian infrastructure? [14:31] How do you do all the things that are so susceptible to human error under the time crunch of what's [14:36] doing in a way that has much higher repeatability and accuracy so you can get more throughput to the [14:41] system? Well, that goes to a broader argument about the elimination of some of the people who were [14:46] involved in mitigating civilian impact from strikes. Uh, that was a Pentagon decision by Secretary [14:52] Hegseth. Um, in the case, you just talked about the thousands of strikes and there were so many [14:58] successful strikes in terms of the targeting in Iran, but there was also, as you know, this horrific one [15:03] in Minab that killed at least 123 school children. Um, Bloomberg reported some personnel relied too much [15:11] on the AI in the Maven system made by Palantir using AI enabled software. Have you determined what [15:19] happened there? Well, we should wait for the final report, but what I can say is that reporting is [15:23] inaccurate. The reporting from Bloomberg you're saying is inaccurate? Yeah, it's inaccurate. Uh, [15:27] and, and actually, uh, the likely cause is human error. And then in quickly following that, [15:33] we actually built additional AI agents that did more checking. That's part of the process we don't own. [15:37] Uh, but responding to the urgency of the moment, we built additional AI that actually double checks [15:43] more of the work outside of the process that we own to ensure things like that couldn't happen again. [15:46] So you're suggesting it was outdated intelligence or human error that fed information into the [15:53] computer system? Is that what you're saying? Well, we should wait for the final report, but I think [15:59] Do you know when that'll be released? I don't. Because the initial reports have been widely [16:03] reported on, including by this network. Um, do you see a harm in not being made public soon? [16:09] Or is it helpful to what you say are improvements you want to make? [16:13] Well, I, I will defer to the department on the, on the timing of how, I have no control over that. [16:17] But yeah, I would love the, the truth to get out there. But do you believe that it was human error? [16:22] That's what you're saying? Not what they're reporting, which is that they were relying too much on the AI? [16:28] Yeah. Um, there are 120 House Democrats who wrote to Secretary Hegseth about [16:35] this particular case and the use of AI in the strike, including Maven. Um, do you have any [16:42] intention to go brief Congress about it or to share information? Uh, we have given briefings to Congress [16:47] and, and they can, you know, Congress is well aware of the capabilities of Maven. [16:51] Mm-hmm. But in terms of the strike and what went wrong, obviously something went wrong. [16:57] I would defer to the department on that. That's, that's not our place. [16:59] Mm-hmm. But it is because you were such a, Palantir is such, um, so interwoven with the [17:06] U.S. government. I mean, it's billions of dollars of, of government contracts right now. Uh, do you [17:12] think more information should be briefed to Congress about the benefits along with the risks? [17:18] Sadly, I'm just a technologist. I mean, on, on these policy issues, I would defer to others. [17:22] Um, well, I want to quickly ask you before, uh, we go about the modern [17:27] modernization efforts, which you are part of. I understand you are a lieutenant colonel in the [17:31] Army Reserves in the Army Executive Innovation Corps. Correct. Which is new. Yeah. Um, there are [17:38] also executives from Meta and OpenAI. So what are you working on there? Uh, my principal role is to [17:44] help with Army Human Resourcing in the G1 department here. So strategies on, um, how do we, you know, [17:49] the Army is drowning in talent, what we would call green suitors, who actually know how to code, [17:54] and they're, they're, they're exceptional. Um, how do we employ them? You know, they don't fit in [17:59] the kind of industrial era model, the Napoleonic codes. Their skills are kind of rank blind. You [18:05] might have an E4 who's an exceptionally good coder. There's not an obvious career path for them within [18:09] the Army. The Army doesn't know how to manage that talent. So how do we identify who these exceptional [18:13] people are? How do we protect them? How do we put them on projects that fully utilize their skills, [18:18] uh, and, and turn them into valuable assets for, for the military? Mm-hmm. So that's your role [18:23] within this corps? That's right. So for people who look at that and say these big corporations [18:28] could have a conflict of interest because they're also doing business with the U.S. government, [18:32] how do you respond to that? Well, it's carefully managed by, by the, [18:36] the lawyers in the department. We're only allowed to work on projects where we don't have conflicts. [18:40] Mm-hmm. And do you think... And I would just say more broadly, I mean, [18:43] it is the greatest honor to be able to serve in our U.S. military. You know, [18:46] as, as an immigrant to this country, it's one of these opportunities. My father, my late father, [18:49] would have loved to have seen me go to a service academy. Life worked out the way it worked out. [18:53] Arguably, I hope I have more to give back to my country in this capacity at 44 than I did at 24. [18:59] And, um, it, it's a great pride. Thank you for your time today. Thank you. I am.