マンハッタン計画、再び：連邦政府のAI活用とDOEのGenesis Mission

We're here today with Barnett Coven, MD of AI and Data at Accenture Federal Services, talking about the impact of data on driving evidence-based decision-making in the federal government. We're talking about how the progress of AI is accelerating that trajectory, as well as projects such as the Department of Energy's Genesis mission and what AFS is doing to support that. Thanks to AFS for sponsoring this podcast, and welcome, Barnett, to China Talk. Thank you so much for having me. It's a pleasure to be here. Let's do a bit of the career arc. So how do we get from a Ph.D. on coin to DOE and the Genesis mission? Yeah, absolutely. So I spent a lot of time in Latin America, sort of saw the after effects of some virulent insurgencies in Peru and Colombia in particular. Ended up doing a Ph.D. studying sort of the spatiotemporal diffusion of insurgent violence as a result of counterinsurgency strategy,

specifically as a result of development assistance and other sort of non-kinetic interventions during counterinsurgency. Really loved that. Sought out the opportunity to do more policy-relevant research, which is not necessarily the norm in the academy. Ended up in a faculty job at the University of Maryland, which was also wonderful. but at a sort of research institution where I was encouraged and sort of able to do a bit more government-funded research, as well as sort of the traditional sort of teaching and publishing and what you'd expect from a professor, and enjoyed that immensely. However, I did a lot of mixed methods work, and it became sort of harder and harder to do the travel that I wanted to do and do the data collection, at which point I had a number of friends who were in a sort of competitive sort of consulting firm firm who had reached out to me over the years and tried to get me to sort of come join them. I finally sort of said, you know what, what the heck, let me at least have the conversation and realize through some really influential leaders at that firm that this would be an

opportunity to apply sort of the rigorous research methodology and the quantitative skills that I learned in grad school and honed at UMD to sort of real world problems and do it in a way that felt a bit sort of closer to the mission and closer to the policymakers in government who were actually sort of hopefully benefiting and using those insights. So I did that for a number of years. I met Garrett at this other firm. He sort of very quickly thereafter went to government, where he became my client for a number of years. We worked on some really cool projects at the State Department, some of which focused on China. Or some of the things you threw your way. Yeah, absolutely. And so in general terms, sort of two of the coolest projects were using natural language processing in order to extract insights from diplomatic cables. And so you may not know this, but the Department of State sends back on order of 3,000 cables between sort of post and sort of headquarters in D.C. every day. These are chock full of sort of insights in textual form, but no one's reading 3,000 cables,

let alone sort of memorizing them and absorbing all of that on a daily basis. So the goal here was to use sort of old school, sort of pre-Gen AI revolution natural language processing approaches to try and extract and categorize some of that insight to make it easily searchable so that it could inform sort of policymaking. And I'm very, very proud of the work we had done there. That being said, if you fast forward half a decade to today, I would have done it completely differently. And sort of some of the tools that are now available that weren't then would have accelerated what was a sort of six-month fast-paced project involving a sort of half-dozen-person team into something that two people could have done. in a matter of weeks and just as well or better. Yeah, well, let's talk about that for a second. The idea that, like, I don't know, software is really cheap, like anyone can, you know, cloud code, codex stuff. In your interactions with the government, like, to what extent has the light bulb gone off

that just, like, a lot more is possible than even was, I don't know, 12 months ago? I mean, the cop-out answer here is it depends. And certain government customers are sort of way closer to the forefront than others. And part of this is a sort of product of their unique mission and how applicable it is to these changes. But part of it is also just sort of risk aversion. Some of our law enforcement clients, for example, tend to exhibit a little bit more risk aversion here. That being said, I think some of the skepticism is really healthy. Wait, sorry, law as in like I can't prove a case if an LLM collects the evidence for me or like what do you mean by that? So I just meant sort of nervous about sort of second and third order implications, whether that be sort of sensitive information is exposed to the public or the evidentiary value is diminished because of a sort of court ruling. And there have been court rulings around sort of Jenny I produce sort of legal briefs and what have you. So I didn't mean anything specific, but I think there's a number of different directions to take that. Gotcha. Okay, sorry.

But the ones who are the believers, like what's the common line there? So I think there's a couple of explanations there. Some of the true believers are coming from industry into government for the first time, and they've sort of seen the power of these tools in their prior jobs in corporate America, where the uptick has maybe occurred first. Others are just super passionate about their mission, and while they've been in government for perhaps decades in some cases, they make it a point to stay abreast of the latest and greatest and are willing to try just about anything if it portends significant advances in their sort of niche domain, in their sort of scope of work. Yeah, I mean, it's an interesting dynamic that we've seen with these labs, right? Because on the one hand, like, I can come up with, like, crazy fun little applications. And, like, there are power users who are, you know, building billion-dollar businesses with a head count of three. But you also have these partnerships that all these labs are signing with consulting firms, with PE firms, because like, no, you can't just press a button and have the AI solve everything for you today.

Like they're clearly, you know, as of May 2026, like needs to be this human layer of actually people to like see the problems and like solve them. I just found that to be interesting that, like, we're not quite here yet where you can, like, all right, let, you know, I mean, maybe, like, GPT 7.0 or Claude Mythos 4.0 can just, like, solve everything in government. But it's not quite there yet. I'm fairly skeptical that we're going to get there anytime soon. And I think there's a couple of dynamics at play here. First, these models are trained on what's available on the open Internet, which is often in many ways distinct from sort of the niche missions of corporate customers as well as government, all of whom have very sensitive data that is not usually available sort of in the public domain and available for training of these tools. So to think these tools out of the box can do sort of everyday tasks that you and I do, like plan a grocery risk, absolutely. That doesn't necessarily, it's not a perfect parallel to sort of engaging in complex foreign policy negotiations per se.

But then the second piece that I think is critical is the garbage in, garbage out problem that's always been the case of data science is exacerbated here. And someone needs to have the sort of, not only the institutional knowledge, but also sort of know where the data lives, what limitations are evident in that sort of extant data, and so on and so forth, in order to get the best value out of those models. And the final thing I'll say is metadata management has become sort of huge here. Having sort of clean, curated metadata gives an LLM some additional context to grab onto that massively improves the quality of the output. You mentioned, like, AI for high-stakes foreign policy negotiations. One of the things I've been thinking about recently is, like, what an eval would look like for, like, a, I mean, you know, tactical level decisions you can kind of understand that you used to do coin stuff. Like, all right, if you give a model a whole lot of cameras and, like, over time they can, like, figure out who's, like, leaving the house at weird hours, like, un-gone on their show.

but if you're talking about foreign policy decision making what do you want to ensure is in the Ukraine peace deal the scary thing is these models are so good and they seem like they can make your life so much easier and they can write your Washington Post op-ed for you the temptation to help think through, okay, Putin gave me these 10 points what are the 10 points I want to throw back is going to be really high. But I wonder if you have any thoughts on, like, what, you know, like, what is the strategic level of negotiation e-value you would want to throw at these models in order to, like, actually have confidence? So I think the easy answer is compare it against the humans. Give the sort of Russia desk officer and the model the exact same prompts and see what they come back with. And sort of compare across the results. And if this sort of model is coming back with sort of similar conclusions to the Russia desk officer, one of two things is happening.

They're both performing equally well or they're both biased by the same things and performing equally poorly. And this sort of phenomenon of mirror imaging, right? We pay people in the U.S. government all the time to go really deep on Russia, really deep on China, whoever the adversary de jure is. And yet they always sort of do so from sort of their own sort of perspective as a U.S. military officer and assumed at least subconsciously that the system works the same. In that sort of, quote-unquote, competitor or adversary nation, and I think that's problematic. The alternative, though, is when they diverge, then you can start interrogating why they diverge, and if you're smart and you have the background information, you can sort of trip up the model and you can see where it's making mistakes, or you can say, uh-huh, I actually learned something here, and the reasoning that was involved is really novel, and we didn't think of that, and we need to incorporate that. Yeah, no, it's interesting. I was actually having a conversation this morning with someone who was in the deep state for a long time and now is on the outside and gets to hire an international team. And so has a, you know, a Japanese national living in Tokyo as their Japan analyst.

And they're just like, look, you know, it was a lot of fun, but I worked with Americans exclusively in my entire, you know, my entire inbox was full of Americans and their analysis of the world. So this idea of these models being able to, like, you know, inhabit different hats or different perspectives, which are not like, you know, U.S. citizen living in 2026 is an interesting one. Though, I mean, if you're only using American models, is there something you're missing? Or if we're trained on the, you know, by and large English language Internet, like there's. That would be a fun thing to tune, actually. It just like okay pretend you you know a China desk officer you know like you know an MSS analyst or Australian analyst Like how do you think this stuff through All right One more coin thing before we jump into this stuff Like, I don't know. Is AI going to, can AI help you win an insurgency besides just doing targeting? And we weren't, like, bad at targeting, right? That wasn't the problem.

I think the issue there was sort of being able to separate sort of the insurgents from the sort of general population, right? If you could identify who the insurgents were, sort of figuring out a solution, whether that was kinetic or otherwise, was not particularly challenging. You're right. I think it's a really interesting question. And I think one of the reasons why the U.S. did reasonably well at counterinsurgency was the willingness of the U.S. military to devolve decision-making authority down to junior officers and non-commissioned officers. that works because of the sort of high sort of educational background and sort of level of training in the force and the sort of generalized degree of trust. That being said, having sort of machine learning models available on the edge on sort of tactical devices, I think, could help with additional context and perhaps also be a sounding board for those JOs and NCOs as they're being asked to make incredibly difficult decisions rapidly with very limited context.

Yeah, I mean, the thing about the, like, Russian negotiation with Putin to end the Ukraine war is, like, that's a one of one, right? And you can't run an experiment about it. But, like, okay, I have 1,000 villages and, like, 10,000 different officers running around, and someone's going to, you know, build the water treatment plant first. Someone's going to invest in the police station. And, like, you know, you can learn from that over time in a given situation. our given time period in a way in which I think the sort of, like, actual, like, heads of state negotiation stuff is just going to be incredibly difficult. But, okay, I'm a believer. Not to say we should, like, just start a word and test it out, but, you know. So let's come to some of the work that you guys have been doing with the federal government since you headed over to Accenture. Start with commerce, maybe? Like, what is the dream of how data and AI can unlock better export controls enforcement.

Yeah, at least as it relates to your podcast in China, I think this sort of topic de jure is export control policy. And what we've seen, I think, there's sort of two sides to this coin. One, what should actually be controlled? And then two, how effective are those controls? Noting that whether it's China or another state, you have a thinking opponent who is going to constantly sort of improve upon the ways in which they try and skirt those regulations and obtain whatever technology it is they're seeking. And in this current moment, we're largely talking sort of advanced semiconductors and to a lesser extent sort of critical minerals, although I think China has done quite a good job at cornering the market on much of that latter supply chain. And here I think there's a couple things we can do. One, some predictive modeling based on past behavior to try and sort of identify higher risk entities. And the Office of Foreign Asset Control and the Treasury has done this for quite some time with respect to sort of like drug dealers and others and trying to identify sort of what accounts to sanction next.

The same logic could apply towards that targeting. But the other aspect is unwrapping this sort of maze of shell companies. and first off, licit companies, especially sort of Fortune 500 companies, have multiple subsidiaries for various legitimate legal and tax purposes. So shell companies in and of themselves are both very common, but not necessarily indicative of nefarious behavior. And so one thing that we found really helpful is using entity analytics, so like entity resolution, in order to try and make informed decisions about how to merge different publicly available as well as sort of proprietary U.S. government data sets, but think like Bloomberg and S&P corporate records data, in order to create a sort of master repository of sort of the companies and their relationships to each other to make it easier to identify that, yes, in fact, this sort of shell company that we've never heard of before is in fact through seven other sort of corporate entities connected to a procurement arm of the PLA.

Yeah, and is that, Like there are edge cases, so now a frontier model can like do the work that like someone would have had to like stare at the two things and be like, oh, this is the same company now. Is that how this is getting accelerated these days? That's correct. And you don't necessarily need Gen AI for this, right? Sort of more traditional AI ML can also be used. There's sort of various sort of nuanced questions that you would want to ask in order to determine whether or not sort of an LLM or a frontier model would actually make sense here. But in general, yes. I mean, I guess the question is, like, is, you know, what is the ratio of stuff where just, like, the government, you know, you mentioned you guys have a Databricks partnership. What percentage of this stuff is just, like, level one, like, this should have been done in 2016? This is making, but sort of the new technology we have today is just, like, making it cheaper and faster versus, like, oh, man, here's all this, like, you know,

We're catching people we never would have caught if it wasn't for the technology that's available today. Yeah, so historically I would have told you, like sort of pre the last four or five years in the Gen AI revolution, I would have told you that 80% of the work that a high-end sort of data science team is actually sort of relatively run-of-the-mill data engineering so that sort of cleaning and transforming of one's underlying data in order to enable the advanced analytics. I think a ton of that is still happening. The reality is a lot of our data in government but also in commercial just isn't ready for AIML, let alone sort of Gen AI. That being said, so maybe it's gone from 80% to sort of 65% in the intervening five years. That being said, the LLMs themselves can be used for a lot of this rote sort of ETL work as an accelerator, which also means there's more of an impetus for our customers, be they government or commercial, to actually do that annoying, painstaking sort of data cleansing work,

because now I can do it at a third or a fourth the cost of what it was five years ago. So, like, the ambition, like, you can just dream more with your budget living at BIS that, like, I can accomplish a project that's three times as big as I could have three years ago. A hundred percent. I want to talk about talent. So, you know, you guys are a federal contractor. Yes. We've had a lot of discussion on past trying to talk episodes about sort of declining state capacity and the challenges and the sort of the need for a government to be able to, like, I don't know, cook the food that it eats. What's the I don't know. Why should why should AFS exist? What's the justification for the government relying on, you know, outside players to help them upskill in this regard? Yeah, and the answer I'm going to give you, I think, is specific to AFS and just a handful of our competitors that have really robust commercial businesses,

because I think that truly is the differentiator. And specifically, we have about 15,000 employees dedicated solely to serving the U.S. federal government. we have 800,000 employees globally principally serving the global Fortune 3000. And we've all heard sort of these biases that sort of corporate America is ahead or the corporate world in general is ahead of the U.S. government in terms of innovation. That's a whole sort of longer discussion that we could unpack. But the reality is sort of the drivers and the incentive structures are quite different. That being said, if you have 800,000 people serving the global Fortune 3000 who are innovating at RapidClip, the ability to lift and shift those innovations into government, I think, is immensely valuable. Let the U.S. government sort of take the sort of insights that we've sort of developed over the years with our commercial clients and benefit from them from day one, as opposed to having to invest and sort of solve these on their own.

It also just gives us immense reachback. If the government wants to do something new and novel, and they can't hire a person as a Fed to do this because they only need the skill for two weeks, we have that sort of novel person somewhere in our organization that can be brought in and that can be sort of productively used to serve in commercial customers most of the time. And I think that flexibility is also part of the value. I mean, from like a personnel perspective, it's interesting thinking about, especially on the sort of like data engineering, data analysts, broadly like software work side. You know, you could have 10X engineers back in the before times. And now you can have like 100X or 1000X engineers if there's someone who really understands how to burn tokens in a productive way. I mean, we've had like kind of hysterical things coming out of meta where they're like firing people who aren't burning enough tokens. Same thing with our friends at Semi Analysis.

It's like, you know, 50 grand an analyst per quarter or something absurd. You know, as you're sort of interacting with different corners of the government, like, have you seen these people? Do they exist out there? And sort of what are they looking, you know, what are the craziest things you see folks kind of dreaming up and how has that interacted with your work at all? Yeah, so we support sort of DARPA and IARPA. in other parts of sort of the AFS ecosystem. And obviously those are the sort of parts of the government that are sort of doing this sort of really wild experimentation and might have a sort of closer mindset to what you're describing at Meta. And I've heard similar things with actually much higher numbers even in terms of expected tokens consumption. That being said, I think where we ought to be heading, though, for most of the enterprise use cases in sort of government as well as established corporations is going to be sort of how do we get smarter about where we sort of use generative AI and agentic AI.

So that doesn't mean just throwing tokens blindly at the problem, because that doesn't lead to a sustainable sort of enterprise-scale solution. But rather, how can we sort of bring in sort of the traditional non-generative AI ML that has been tried and true for decades, have that be 80 or 90 percent of a solution because it works and works well, and then you sprinkle on sort of gen and agentic AI on top for sort of the niche aspects of the problem where sort of traditional AI ML is not going to be able to perform as well. And in doing so, you constrain token consumption, so you keep the total cost of ownership down, but you actually get better results by virtue of using sort of deterministic processes as opposed to probabilistic ones for pieces of the problem that can be answered deterministically And I think that's the special sauce, and that's what a firm like AFS can bring to the table. That doesn't mean I'm not all for throwing everything at an LLM when we're in the really early experimental phases,

way before we consider sort of moving something to production. And that, frankly, does allow our teams as well as others to just be much more innovative and try things and fail, but do so faster and then learn from that failure. Yeah, I mean, you see CEOs like now having to take into account, like I'm not just hiring an engineer. I'm hiring an engineer plus like whatever, $10,000, $100,000, $500,000 of like CapEx that just like goes along with this human being. And like the answer is we don't know, right? And it's going to change over time. And I think the sort of like is the federal budget process going to, you know, like like you get head counts oftentimes from Congress. Right. Not necessarily like compute budgets. And just speaking of BIS, like there was this like multi year long effort to like try to get them 100 million more dollars. and, you know, if we're sort of, like, focused on executing the mission, right?

Like, is the mission 50 more case officers? Is the mission five case officers? And, like, you know, a ton of models just, like, spun up, tracking every news story and every ship on the planet to, like, figure out what's getting smuggled somewhere. I don't know. It's the sort of, like, what is, like, the optimal way to spend government money to get the outcome that you're looking for is probably, like, the most wide open that it's been in a really long time. Absolutely. And so the government budgeting process, realistically, your budget this year was decided two years ago, give or take. And that is problematic here. And it's not just the headcount that you were mentioning, but also what's your token consumption? What's your underlying compute bill from your cloud service provider, right? That is going to become harder and harder to predict, given the sort of rapid pace of innovation, the rapid pace of change. I think the ways to sort of solve this while still providing accountability are, one, provide a bit more fungibility sort of between tokens for this model versus tokens for that model,

tokens for this use case versus tokens for that use case, for example, as long as they're sort of tied to overarching priorities, and also to aggregate across a department or agency. because it may very well be the case that this particular office at a given agency has a really novel use case that we couldn't have predicted. But, you know what, this other sort of department over there has a problem that's going to be much harder to simply solve away by throwing tokens at the problem and having some fungibility within sort of the overall department in terms of how that budget gets spent with appropriate sort of thin-offs in place so that there is accountability, right? I am not suggesting we blindly spend taxpayer dollars or just remove constraints here, is going to be sort of the direction the government has to go if they want to continue to sort of leverage these sort of latest and greatest innovations. Do you want to talk a little bit about the kind of, like, national security piece of this? Because, you know, if you want to do, like, all this stuff on a clear data center or whatever,

that's a whole other lead time, which isn't just, okay, you know, shifting it from BIS to the CHIPS Act or whatever. and commerce, but, like, you've got to build a place and buy the chips, which is not changing a line in a budget. Yeah, absolutely. And so I had conversations in the last week with two different foundation model companies, both of whom were sort of touting that, hey, we are releasing the latest and greatest models into our sort of FedRAMP but unclassified U.S. government environment the same day they're becoming available to the public. The reality is once you sort of get to sort of IL 6 plus to those classified environments, that is not the case. And that realistically can't be the case given sort of security constraints here. That being said, it is very much an arms race. I had this amazing Ph.D. student who was an Army cyber officer who sort of came back for a Ph.D. And she sort of she looked at sort of risk aversion in the A.I. era in government.

and her sort of most powerful insight was inaction is in and of itself risky, right? We so often focus about sort of buying down every risk for the action we're going to take and we potentially don't take that action because of the risk we can see that we forget about the fact that if we don't sort of do something, there is not only a risk but a guarantee that we will be left behind by sort of peer and near-peer competitors who are less risk-averse than us and are willing to sort of take risk after risk after risk. And I do think we've seen that with China in particular. Can you talk a little bit about, like, what it takes from bringing something from an unclassified to a classified network? I mean, you're just, like, worried Mythos is going to run around and, like, delete stuff? Or is it just the compute isn't there in the secret data center? I mean, so the compute is actually less of a problem in classified environments. It's more sort of what guardrails need to be in place. So is there any chance that any of this data could get out into the wild?

And again, harder in a classified environment, but not impossible. Is there a risk that this model is going to sort of inject something into our environment that could be deleterious? And frankly, has the model been poisoned in such a way that it's going to give us the wrong answers? And there's been some amazing research on sort of poisoning, okay, which you're clearly tracking based on reaction there. It's not particularly hard, and if I were sort of in MSS or 2PLA, it'd be fairly easy for me to figure out sort of what my counterpart at CIA or in one of the sort of military intelligence organizations might want to do with an LLM and to get ahead of that. Unfortunately, this is a bit of a gun armor race, where in the cost of the sort of offense increases linearly over time, where the cost of sort of the defense increases exponentially, i.e. there's an inherent advantage to someone who's sort of trying to do bad when we're talking sort of classified spaces, and so there's just a need to be especially cautious there. Let's talk a little about the Department of Energy.

What is the Genesis mission? Yeah, absolutely. So the Department of Energy announced Genesis mission, And then there was this signing ceremony at the White House that our CEO, Ron Ash, participated in. We are in originally 26 different companies, so Accenture Federal, and then mostly sort of foundation model companies, the sort of hyperscalers, et cetera, got together and signed this sort of commitment to work with DOE and work with the U.S. government in order to deliver on this promise. And the promise is, in effect, the DOE will become sort of a leading organization across the U.S. government for generative AI and that we're going to treat sort of advances in Gen.A.I. sort of similar, the way I've heard it described from some of our clients, is similar to the Manhattan Project, which is convenient since you do have that sort of poster back there for the 21st century. Provide the appropriate resourcing in order to make sure that the U.S. sort of remains preeminent in this domain and, importantly, ensure that we can safely sort of build a moat around sort of critical sort of scientific discovery

and other use cases to de-risk the application at Gen AI. Because my point to you a moment ago was not that we shouldn't use these tools in a classified environment, just that more care is needed. In fact, not using them is the wrong answer because, again, to my sort of risk comment. All right, great. So let's start with how broken the scientific ecosystem is today from a data sharing perspective. Absolutely. And let me sort of take this back to my sort of academic sort of backgrounds here. I think one of the things that really irked me about academic publishing, and so as you probably know, sort of professors are evaluated more or less on research grants and publications, and in particular publications in top-tier journals. For those of you who are paying for college educations, teaching effectiveness factors in way more in sort of like liberal arts colleges than sort of R1 universities, although that is still a consideration. But with that said, one of my sort of big quips with public academic publishing was you only ever really got published if you sort of generated novel insights that were statistically significant.

And the problem is this sort of research enterprise is one of trial and error, right? For everything I published, I tried 10 or 15 things that didn't work. And when I say didn't work, they didn't work in the sense that I didn't provide, didn't sort of have this aha moment. It didn't mean I didn't learn something in the process. Let's do an example. Let's give us a coin case study here. Ooh, okay. So a lot of the original, and this gets to my dissertation project, but a lot of the original research in counterinsurgency that looked at the use of development aid and governance reform, looked at the use of development assistance in a given village or area on sort of insurgent violence in the village, and completely ignored the fact that in a sort of insurgency, what's happening in one village is observable to the insurgent as well as to sort of residents in other areas. So I tried to sort of fix this by implementing sort of a spatially weighted regression approach to studying this

so that we could measure the effect of a development aid intervention not only on the village receiving it, but on violence in surrounding areas. And what I ended up finding was that on average, you actually only serve to displace and possibly exacerbate the violence. So in sort of this original published research, you said, yes, development aid is good and it reduces violence, when in reality the opposite was true. That being said, a lot of the early studies, including some of my own early research to try and prove this, had sort of negligible effect sizes and barely achieved statistical significance. I had to sort of go back to the medical research and pull in models that were used actually in sort of cancer research for modeling sort of the diffusion of sort of cancer or the sort of reduction in sort of the spread of cancer as a result of chemotherapy and other sort of clinical interventions, and lift and shift that to a totally different domain to sort of find the appropriate approach that worked. But in doing so and in the failures to achieve statistical significance

prior to finding these sort of appropriate models and applying them to a novel solution and a novel domain, none of that research was out there. And the U.S. government was providing ample amounts of assistance without considering these lessons, potentially at the cost of sort of not only treasure, but also sort of the blood of US service members. And that's unacceptable. And to which there needs to be a way to make sure that those failures that contain valuable information even if it not statistically significant and proven beyond to a 95 confidence interval or whatever the threshold is in one particular discipline but that was it for me need to be out there and discoverable to other researchers If that's not the case, at worst, we have 10 people spending sort of resources on the same work without even realizing it and without sort of benefiting from the insights of the other. Sorry, that's at best. At worst, some of these important insights are lost forever. So let's now do like a Department of Energy Genesis case study.

Yeah. So my critique there is focused largely on sort of traditional academic research, sort of typically something that's done as a solo sport, one professor or grad student by him or herself working on their own project in a silo. The way in which DOE and the national labs do sort of experimentation and research is a little bit different and a little bit more collaborative in nature. That being said, there are still important silos, and it's not particularly easy to share data across the labs, all of which have their own sort of distinct tech stacks, as well as to sort of share that information across sort of teams within industry and academia. And so the first step is to sort of enable that sort of data sharing and enable a common pool of understanding. But the second step is to have those failures be accessible and mineable, whether it's by a large language model, a human or some other approach for novel insights. Because, again, the sort of nagging question for me is how much value that we've already invested taxpayer resources on is sort of laying at the cutting room on the cutting room floor never to be found again.

Yeah. And so, I don't know, how's AFS going to solve the problem here? So what we're doing is we're building what we've sort of colloquially termed the backbone for scientific research. And this is the infrastructure that will allow secure collaboration across the labs, DOE headquarters, NNSA, critical sort of industry and academic partners with the tooling to pull in data from sort of commercially available sort of instrumentation that you would find in sort of hard sciences research labs, sort of equipment made by the likes of Siemens and what have you, and sort of make it accessible with robust access controls, but accessible to those who have sort of a need to know and for which it would be valuable to their research enterprise. The hard part now is not so much sort of coming up with the latest and greatest sort of foundational model. And frankly, their company, the foundation model companies that are doing amazing at that. It's how do you connect those models to a niche use case like sort of scientific research

and ensure they have access to the relevant training data and sort of the relevant insights? What is like a thing that Sandia would be doing today that like not enough people would be aware of or like an insight that would have disappeared without this in 2028? Yes, I am not a material scientist, so take this with a grain of salt. But I think one answer would be around sort of supply chain security. And so one aspect of supply chain security is where are there opportunities to sort of change the way in which we produce critical minerals. So beryllium is a great example. It can be actually made from a number of different types of ore, some of which have not been sort of commercially viable in the past but maybe now, how much of that data was discarded 100 years ago because there was a cheaper, easier way to do it at the time that is now no longer possible because a strategic competitor has cornered the market for that sort of mineral type. And then where is the data on what happens to exist buried in the ground in Utah and Nevada and what have you

that might be sort of relevant to solving this problem? and sort of where it's sort of separately to those sort of geological studies, where is the sort of underlying data on sort of novel approaches to adjusting how we treat sort of tailings from existing sort of mining operations in order to try and recover things that in the past were sort of just waste minerals because, again, it didn't make sense or we didn't have the technology to actually extract them. So one of the critiques you've or one of the reasons like this sort of like the exhaust fumes of scientific research hasn't gotten hasn't gotten out is not just that, like, you can only public publish results with statistical significance, but also like it's annoying and it's more work for the researchers themselves. to, like, okay, if I only have to write up 12 things, now I have to, like, upload 11 other data sets. How are you, you know, what's the solution on that front?

And AI can write the reports for you. And, I mean, I'm being a little bit glib here, and I do think there still needs to be a human in the loop here, but if it can do 80% of the work, why not? And we're seeing this within sort of data engineering. One thing that all of the data engineers, almost to a fault that I supervise, sort of hate doing is documenting their code. But it's critically important so that the next person can sort of pick off where they left off, and that's so we can interrogate. Did they actually do what they think they did, or is there a mistake here that's potentially consequential? We've seen amazing strides from sort of Claude Code and sort of Codex in terms of sort of doing that sort of commenting out or documentation of Code for us. The same logic applies really nicely to report writing, and frankly there's an amazing corpus of text in terms of published peer-reviewed sort of articles for what that should look like, such that it wouldn't be unreasonable to expect that a model could get you 80% of the way there. And since you could even sort of conceive of a paradigm where this is sort of a secondary publication, a publication of failures, which I would absolutely read,

wherein maybe the ethical standards are relaxed, where we allow an LLM to have primary authorship. But if I'm doing physical experiments, right, like it's important that, I don't know, I do all the steps right and remember all the steps right. I mean, is the actual end state is I just have like a camera in my lab that's like taking notes and writing up everything I do over time. So if I only pour, you know, half a beaker in instead of a beaker, like it'll notice and write that up and upload it to your thing as well as like let me know that like there's some user error in here, which is why I didn't find the, you know, new beryllium compound. Yeah, and it could be that. I mean, I think in reality, a lot of this research is a little bit more, research is a little bit more automated and whatnot, so it's not me hand-measuring, but rather using instrumentation to do that, to which the instrumentation is producing output that, by and large, is not being sort of captured or at least not retained. So I don't even think we have to install the physical cameras, but we just need to collect that sort of data and actually do something with it.

Cool. I'm excited for that. I think this is, like, a net positive for humanity. We're going to, like, find more cool stuff. On the, like, can you talk about the, like, secured versus, like, not secured part of this work? Yeah, absolutely. So one of the key things to make this palatable for research when it comes to critical minerals, nuclear energy, grid reliability, right, all sort of aspects of, frankly, our competition with peer and near-peer competitors that are going to be consequential, is making sure that while this data is easily shareable amongst those we want to have access, there are robust gargrails in place. So these should be models that exist on existing sort of secure sort of cloud environments or even perhaps bare metal. Our national labs have made major investments in sort of high-performance compute, so they have the fair metal infrastructure that most sort of corporations and others do not. But so sort of deploying models in secure spaces, but also implementing really robust

RBAC and ABAC controls such that you can be extremely granular down to the row or column level who should have access under what circumstances for what research. And it would likely be the case that you'd want to share more widely within the sort of national labs than you would with certain university partners, than you would with certain international university partners, and so on and so forth. So that granularity is going to be critical to making this enterprise work. Roles-based and attribute-based access controls. So the idea here in lay terms is to make it really, really easy to decide that you, based on your sort of specific role within an organization, should have access. And so on government systems as well as most corporate systems, you are signing in with some type of either physical token or unique sort of login and password and whatnot. So using that single sign on, it will identify that you are this person. This is your role within the larger research enterprise. These are the projects you're working on and so on and so forth and automatically govern access. Importantly, according to the rules set in place by the individuals or organizations that generated that data.

And I think one critical piece to this beyond the technological is how do we get sort of the buy-in from the various stakeholders? And the critical piece there is going to be ensuring that ultimately they have discretion over how the data they generate is shared and can sort of impart that sort of knowledge. But then these RBAC and ABAC controls mean that you don't have to go e-mail John over there and say, here's what I'm working on. Can I please have access? And then John doesn't go from being a full-time research scientist to a full-time e-mail answerer and permissions grantor. But if I'm the Secretary of Energy and I'm cutting all the checks anyways, can I just say, sorry, you have to share this with everyone I think is responsible? Like, who's, like, what's the actual friction point? Yeah, so in that case, the answer is nuanced. And I will caveat, I am not a lawyer. I am the kid of two lawyers, but that only gets me so far. And so when you sort of sign a research grant or contract with the federal government, there are terms and conditions that you agree to. And depending on those terms, that will govern how the information is being used.

Typically, though not exclusively, the government will retain rights to the research they funded for their own purposes. But that doesn't necessarily mean they can go ahead and share that with a third-party sort of commercial entity or sort of university partner or what have you. And again, it's going to be very dependent on the terms and conditions of an individual grant or contract. Are the labs excited about this? I don't know. It must be fun if I'm a scientist. It's like, OK, if I'm only capturing 10 percent of my work now, I'm like just like now you're a 10x scientist and you don't have to do anything else except to like upload every once in a while. I think so. I mean, at this point, while we have sort of viable demos and whatnot, we're not at the point where we're fielding this and they can get hands on. So I think the excitement will really come the second people can get in and play, especially for scientists or just sort of people like me who are sort of naturally inquisitive. the ability to actually touch it and play with it is sort of the real differentiator. Like, I'm pretty excited about the concept, but I'm doubly excited because I've actually played

with what we've been prototyping at AFS. Gotcha. Cool. Anything else you should cover? No, I think that's great. I think this was really fun.