Government Agents: AI Agents Meet Tough Regulations — Mark Myshatyn, Los Alamos National Laboratory

Talk transcript

1. 0:00

   [on hold music] All right, good morning. Uh, my name is Mark Myshatyn.

2. 0:15

   I'm our enterprise AI architect at Los Alamos National Laboratory. Uh, today, you know, this is an AI conference. What, what's a nuclear science lab doing here? The reality is we've actually been doing applied AI ML for almost seventy years.

3. 0:29

   Uh, this is actually one of our scientists in 1956, uh, playing Los Alamos chess, uh, in front of one of our first supercomputers, MANIAC I. And what's unique about this is we-- i-if you look at it, there's actually no bishops on the chess board.

4. 0:44

   You know, we've, we've been doing applied statistics and applied machine learning since we didn't have the memory needed to hold an entire chess board in a computer at once.

5. 0:53

   A-and that's fascinating to, to say. Back when this photo was taken, it was after the Manhattan Project. We were pushing the edge of developing Monte Carlo methods that we still use today.

6. 1:04

   And for us, you know, AI didn't come as a complete surprise, but the opportunity that's come with agents, with things we can do has been incredible, even to, to us that have been along the ri-- the ride for quite some time.

7. 1:20

   Uh, let's see if I can make this full screen.

8. 1:24

   So what we have going on here, this is actually a demonstration. You can find it on our YouTube channel if you can't see it on the screen here. But we've looked at generative AI not only from a strict model standpoint, but also from an agentic standpoint as a way for us to move science faster.

9. 1:43

   Uh, we, we, like much of the federal government, are under a squeeze to do better, faster, cheaper, and more to protect our country. And in this case, going from not just what a model knows, but what we can let a model know was really the, the change that happened here.

10. 2:00

    We started with a problem of go design a, an ICF, an, an inertial confinement fusion capsule for our sister lab at, uh, Livermore, uh, across the bay here. And we said, "Read a paper.

11. 2:12

    Go read lots of papers that you think are tangential to this first paper, and then come up with a design for a fusion capsule." Uh, it created a hypothesis, and the thing that's kind of uniquely ours is this isn't a generic, you know, chatbot that spits back a bunch of code.

12. 2:28

    What you'll see here in a second, we're actually executing that code on our high-performance computing assets, and we are actually running, you know, thermodynamic and hydrodynamic tests on some of these types of problems, where our model, you know, isn't just a, an LLM.

13. 2:45

    It's all of the f-- you know, fifty, sixty plus years of math and science that we've done to, to bring the, the management and the development of our nuclear stockpile and stewardship of that stockpile, bring those tools into an agentic era.

14. 3:00

    So we're looking at this as a chance, uh, for agents to move faster, uh, and for, for science to move faster, because the risk at the same time is starting to move faster.

15. 3:10

    And you can see here, it actually did come up with a design that it thought optimized that yield, and we were simulating a, a slice through an ICF capsule.

16. 3:19

    But okay, that's one nice toy problem. What does that mean for the other twenty thousand researchers that we have at our laboratory? For those of you not familiar, we're forty square miles of labs, test sites, uh, test plants.

17. 3:35

    We have thirteen nuclear facilities, and so we're huge. We have a huge breadth of what we're trying to accomplish with AI, uh, and getting our mission moving faster. For our National Security AI Office, you know, what you just saw, that's the first thing of-- that we're charged with.

18. 3:50

    Push the science of AI faster. Don't just sit there and consume commercial tools or open source tools. We write our stuff. We write our own models. Uh, we also realize that we can't do everything.

19. 4:01

    We don't have the hubris to understand or to say here, "Oh, we understand everything. We don't need anyone's help." We absolutely need those partnerships from commercial industry, from academia.

20. 4:11

    And then just like the rest of you all here, we're looking at how do we bring AI and GenAI tools into our workflows. You know, we have a huge footprint.

21. 4:20

    We have to do payroll. We have to do procurement. We have to do cybersecurity. And so our office is kind of in there, how do we do that? And it really does come down some, uh, to some of what we're doing with our partners.

22. 4:32

    We have some great academic partners. We couldn't-- Uh, at the time these slides were released for, uh, public review, we didn't get the screenshot on there. We also announced a partnership with the UC family of schools, uh, on the academic side of developing, you know, the future of AI.

23. 4:47

    But we're also working with all the frontier labs. You know, here's a couple press releases where we've actually done chem biosafety work with OpenAI, and we, we've been able to acknowledge that work that we've done with them.

24. 4:58

    But we have a place where we've been doing-- We're, we're a safe place to do dangerous things, and we've been doing that for decades. So it's a neat partnership to have these frontier labs that really can afford to hire anyone they want still come to us as a source of data and a source of partnership.

25. 5:14

    There in the middle of that last picture, we actually have a science of AI in the hardware space. Uh, that's our Venado supercomputer. It's over twenty-five hundred nodes of Grace Hopper super chips.

26. 5:26

    And we, we brought it, um, through a partnership with OpenAI or with, um, NVIDIA and, uh, HPE to build a supercomputer that can help us push the boundaries of what does it mean to do AI research.

27. 5:39

    And then more recently, we've also brought OpenAI's models onto this system, brought it up to our classified networks, and we're getting to work on the really hard problems that are unique to our data and our mission space.

28. 5:52

    Uh, when we talk about agents, you know, partnerships take trust. You know, certainly having labs trust you with early access to their models or model weights. As we talk about sharing responsibility with our partners, w- certainly the, the responsibility of what our AI tools and services do starts to matter.

29. 6:09

    Uh, there were-- previous administration had certain executive orders out. Those were replaced largely in January when the new administration took change. But this piece of OMB memorandum just came out in April, uh, M-25-21, and there's M-25-22, and it starts to codify, like, what things should the US government start to worry about when we're fielding these AI systems.

30. 6:32

    It tells the government to go faster. That's important. But it also says these government-type workloads, they have real-world impacts. You know, for us, we are not a T-shirt company.

31. 6:42

    If our data gets out, that's, you know, geopolitical challenges show up, uh, kinetic challenges show up. People can die if we do this wrong. And this, I won't bore you, it's like twenty-five pages, reasonably well-written for an OMB memorandum as far as readability and comprehensiveness.

32. 6:59

    But it says we as the US government need to move faster into bringing this into everything we do. It's not enough to just buy, you know, pick your favorite Office add-in tool and say, "We can type PowerPoints faster or summarize our emails faster."

33. 7:14

    We gotta go deeper into our mission, and that comes with trust. So who here is part of a software-as-a-service, uh, company or startup? Okay, handful of hands here. So you've probably seen something similar to this, especially if you've been in the cloud space recently, that as us as customers start to trust you with our data, your responsibility

34. 7:36

    also comes up. Uh, that's easy to do for our open public unrestricted data, like the open science work like I showed off of our ICF capsule agent. But as we get into controlled and classified, as we get into classified, in the DOE space, as we get into restricted and formerly restricted data, where the physics of how nuclear

35. 7:53

    weapons work don't expire, that, that will forever be classified. It's born classified and stays classified. It takes an element of trust in you all as our builders, as our providers, and this is really some of the most interesting and then frustrating conversations we have with companies trying to sell us tools and services, is, "Great, you have your

36. 8:13

    SOC 2 report." Uh, I have NIST 800-53. This is actually Rev. 4. It's over a thousand, uh, different security controls and enhancements. And the, the US government has put a lot of legislation in place to do traditional cybersecurity work.

37. 8:29

    FedRAMP certainly tried to make this easier by coming in and saying, you know, "Two hundred of your security controls, three hundred, four hundred have been vetted with a third-party authorizer.

38. 8:40

    You have some continuous monitoring." Has anyone here been a downstream of the FedRAMP process? Yeah, I see a couple smiles. So you know how much of a pain this has been, and much like everything else in the government right now, it is changing.

39. 8:53

    There's a new FedRAMP program out there saying, "If we're going to trust you with our data, if we're gonna trust, trust you with the outcomes of our agents, you have to start thinking about your continuous monitoring, your continuous security posture."

40. 9:07

    Uh, if you work with the DoD, that gets even harder. Uh, DoD has what they call their Security Requirements Guide, or CCSRG. Um, it talks about if you're touching this type of data level, so it takes that three types or three types of FedRAMP, it layers on two more, uh, impact levels as the DoD calls them and

41. 9:25

    says, "This is how you're going to access that service if you have PII or mission data or operational data or finance data." And then they add another copy of this book, you know, CNSSI 1253 on top of that.

42. 9:39

    So if, if you're looking at this saying it's a lot of governance, it is. Um, but the fun part is right now, where we are today from those, uh, April third memorandums is

43. 9:53

    AI use cases, AI governance is still on the drawing board. Like, we are in that hundred eighty-day rulemaking period that these, uh, pieces OM-- of OMB memoranda put out saying agent or, uh, agencies have to go develop their strategies, their plans for developing, you know, AI implementations.

44. 10:13

    How do you govern pilots? What's considered high risk, low risk in your context? And there's some prescriptive guidance out there. NIST back in twenty twenty-three released their AI Risk Management Framework, and- Our morning breakout sessions will begin in five minutes.

45. 10:26

    Five minutes for morning breakout sessions. [laughs] Please choose the breakout session of your choice. But the fun part is you can develop the future with your customers right now. You know, th-this is a clean sheet of paper from a technology perspective that- Our morning breakout sessions will begin in five minutes.

46. 10:39

    -we largely haven't had to tackle. Uh, and it's, it's fun in a US government space to say we can invent part of the future together with commercial industry, um, make hopefully better, less obnoxious, less obstructive decisions, so we can keep moving mission faster.

47. 10:56

    A-and if it sounds like this is a lot of lawyers and paperwork, it probably is. Um, there, there's no getting around some of these records and artifacts that do have to exist.

48. 11:05

    But the, the reason you'd wanna collaborate with us is we're doing things that are either incredibly hard or can't be done in commercial industry. Um, at least at Los Alamos, we are sitting on petabytes of data that has never seen the Internet, will never see the Internet.

49. 11:20

    Uh, we have subject matter expertise in chem, uh, bio, materials physics, um, materials composites, um, certainly cybersecurity and the design of high-performance computing that some of the partnerships I mentioned earlier, and they can be your partnerships too.

50. 11:37

    You know, we, we firmly believe that if we're talking about taking care of the country, taking care of our national competitive advantage, that's not just a bunch of scientists sitting on a mountainside in Los Alamos are gonna figure that out.

51. 11:49

    We really do want your help and your, uh, engagement with us to, you know, push the boundaries of what we know. This was originally meant to be an architecture talk, so I'm finishing up with a architecture slide.

52. 12:00

    If you are interested in bringing a-agentic tools, agentic services to the federal government, there's really four things to think about. Now, we want to see that you've built for explainability.

53. 12:10

    Our keynote this morning touched on that a little bit, of how did you get to that decision? You know, if, if something goes wrong or if we have a bad day, we don't have shareholders that we're responsible to.

54. 12:21

    We have the US citizens to be responsible to. Um, we have... Whatever that outcome was that, you know, caused some press briefing, we need to be able to trust our agents the same way we trust our staff.

55. 12:33

    Uh, when we talk about fielding things, again, we, we are not a T-shirt company. Building for isolation matters. And, um, I w-was looking forward to seeing Microsoft's demo on the, uh, uh, uh,

56. 12:45

    self-hosted, uh, AI foundry pieces, but for us, we do that anyways. We look and r-leverage heavily open source tools and services and models to do some of this work because we can't get it from a hyperscaler cloud provider.

57. 12:59

    Uh, so as you're building your tools and services, take a look at some of those services and scope page, even if you are a SaaS startup. Um, if you can build in a DoD Impact Level 5 environment with that limited number of services from your cloud vendor, you can deploy anywhere.

58. 13:14

    You know, you, you have the least common denominator, uh, out of that entire tech stack. If you can deploy your, you know, your tool, your application there, that makes our job easier.

59. 13:24

    That makes you more portable. And as, along with that comes build for governance. We also have some awkward conversations with customers where it's, "Well, we need a software bill of materials as we're doing this procurement with you."

60. 13:37

    And yeah, [laughs] uh, as people look at us like, "I mean, I guess we can dump, you know, what we had in our build script," and it's, it's a little bit of an awkward conversation, but that's required per our regs.

61. 13:49

    You know, the AI stuff is moving a mile a minute. The traditional cybersecurity stuff is moving faster, but not quite there yet. Uh, so if you can plan to have those conversations of how did you handle open source dependencies, what are your patching plans, what...

62. 14:03

    You know, help us fill this paperwork out if we're buying from you as a software as a service or platform as a service. That makes that entire partnership that much faster, that much more friendly.

63. 14:14

    And lastly, keep up the speed. Uh, we have also had some awkward conversations with some of our service providers saying, "Why is your federal stuff a year out of date?"

64. 14:25

    You know, "Why is that service parity not happening a year, three years, five years, uh, from when you launched it in a commercial region?" And that's not us just l-liking ourselves and wanting to have bravado that, "Oh, we're the government.

65. 14:38

    We're, we're a quasi-federal agency. We, we care about our data." No, this is rooted in export compliance law. This is things like we can't buy from you unless you're in the right places.

66. 14:48

    Um, so it's... If you can design for speed in your hard corners, that optimizes your chances of, uh, fielding your tools and services with us, uh, in different places that we have to operate to meet our mission.

67. 15:04

    A-and with that, I mean, Los Alamos, we were founded on the idea that the right application of math and science can change the world overnight. Um, we've, we've done that.

68. 15:14

    We're not a stranger to how that feels to show up and the world is now different. Uh, that's what we were founded to do. And when we look at AI, uh, agentic tools, what we can do with frontier models, a-any of the above, um, we see it as the greatest opportunity and the greatest threat to national security.

69. 15:33

    But the opportunity is what keeps us showing up. We're not scared of the, the downside risk. We have to be here to, to help develop the future. Uh, one of my favorite anecdotes, uh, because we are a nuclear science lab, uh, we do a lot of nuclear non-proliferation work.

70. 15:48

    And because we do that type of work, we've gotten really good at specialty sensors. And what have we been able to do with that specialty sensor? We have a laser strapped to a car on Mars zapping rocks.

71. 15:59

    You know, we built the ChemCam sensor. So even if you're a little bit on the fence about should we engage with, you know, the nuclear enterprise of the US, there's other fundamental science that we do that's just pushing the boundaries that we as a h-human species know and can do and can, can grow into.

72. 16:17

    So, uh, with that, thank you so much for your time today. Uh, really appreciate it, and I'll be available on the side for questions. Thank you. [audience cheering] [upbeat music]