The Exchanges

Q And so what is, what does this enable? Like when this is, let's say you have that football field size quantum computer, what does that enable?

A So the biggest thing that it will enable first, because effectively you can think of it as building molecules in memory and using those molecules. Uh, is going to be material sciences and, uh, chemistry first. So, in fact, one of the targets for Amazon's working backwards document for our quantum computers, a thousand error cryptic qubits could do a Hamiltonian on ammonia. Ammonia is the most produced, uh, we've been producing ammonia since the 19, for almost over, over a hundred years. Um, and it's probably the most produced chemical. It's in fertilizer, it's in petrochemicals, it's in Plastics. It's in just about everything. Um, and it's very expensive and energy intense to produce. We know by watching bacterial interactions that, that it can be produced at low energy state. We just don't know how. So in the past, Like a high temperature superinductor, superinductors in general have been discovered accidentally in the labs and then leveraged, uh, in the future with a Hamiltonian simulation, you can say, here's the outcome I want. Give me the chemical formula that will give it. So you can reverse engineer an outcome in, in chemistry. Um, on today's classical computers for ammonia, if you took all the iPhones and all the laptops and all the Android phones and all the cloud computers on earth, And put that, that simulation into it, it would run for longer than the history of the…

AI assessment note: “biggest thing that it will enable first... is going to be material sciences and, uh, chemistry”

Q Can you define what this, what this is?

A Um, so, um, outcome based firm fixed prices is, you know, to put like, uh, you know, you're going to have a house built. You, you have a price that you pay for the house upfront. Uh, uh, time and material is, you know, you, you, you have a house built and you just pay as you go based on the changes and all those other things. Right. Um, and so, um, both have advantages and both have disadvantages. I think, uh, early days in the government, mo many things were firm fixed price. Um, and outcome based. In other words, you want an outcome at the end. I want to, I want to land a, you know, a person on the moon or I want to do whatever happens to the, you could be an outcome based type contract. Um, sometimes time and material makes a lot of sense when you're asking, uh, the government to do something extraordinary they've never done before. Uh, or no one's ever done like three D printing and organ transplant, right? We don't know that we can do that. No, one's going to sign up for an outcome based contract like that. But, uh, migrating, uh, from on-prem to the cloud should be an outcome-based contract that we know how to do that. So, so when you know how to do something, um, uh, outcome-based makes a lot of sense, uh, and firm fixed price makes a lot of sense. When it's something that the government's really pushing the edge of technology on, that's when you sort of have more of a t…

AI assessment note: “outcome based firm fixed prices is... you have a price that you pay for the house upfront”

Q Okay. I mean, did you see there was that story of, I think it was veterans records being held in a cave? Uh, how does that happen?

A Um, so, so, so that's not really accurate exactly, right? There, there are, there, there is the need, uh, for long-term storage at NARA and other places like that of, of data, uh, that is underground, um, and that data is also stored purposefully, um, uh, on, uh, non-electronic formats, and the reason for that is, Um, we have legal requirements, but the government to keep that data forever. Um, now you could change those legal requirements. There's reasons for those legal requirements. We have legal requirements to keep that data forever. And if you stored it on some type of technology, you'd be constantly having to upgrade that technology. You know, you would have started storing it on, you know, a 1600 DPI tapes. Then you would have had to migrate that to 2800. And then you would have had to migrate that to 62 50. Then you migrate that to 37 K. Then you would have migrate that To disks and then you go and on and on and on and on. Right. Um, so there, there's a certain amount of logic to that. Um, it is stored in OCR character, so it can be, uh, automated at any time. So there, there are, you know, there, there are certain things that are true. Other, other things are, are, are, in my opinion are misrepresented.

AI assessment note: “there is the need, uh, for long-term storage at NARA and other places”

Q this. If we get to a place where AI lives out its promise, uh, what does the public sector, what does it look like? Like, what are the benefits? That we see within the government. Does it enable the government to provide services better? Does it enable us to interact with citizens in a smarter way? Like if we dream about a best case scenario, what does that look like?

A I think that that's exactly where it would be is, is better citizen services, um, a faster, more efficient delivery of citizen services, a reduced overall cost ideally. But remember on the reduced overall cost piece, these models use a lot of GPUs. They are really expensive to train and they are really expensive during inference on today. So that's another area that, that we really question sometimes the ROI of some of these things because of the cost of all of it. Um, so that's another balancing factor. I think we don't have good data yet, um, on the ROI. And so that, that'll be, you know, the cost of operating the model and training the model, um, and running the inference on the model versus the feedback. Um, and I think some of that is we don't have good metrics to be able to track those things. And so we're working on those as well. That is something we're working on, but I would imagine a world that's got better citizen services that can deliver things faster and get things done faster, uh, and do validations faster about it. You know, there's other sides to this, too, where you, you shouldn't go overboard. At some point in time, a citizen should expect to talk to a person.

AI assessment note: “I think that that's exactly where it would be is, is better citizen services”

Q And I, where else do you think it should be used?

A Um, it should be used a lot more, uh, for doing fraud management and financial systems. It could be used a lot more in the IRS. It could be used. I mean, I could go, I mean, there's a lot of other places it can be used to. Um, you know, and, and large language models aren't a panacea. They're not perfect in everything. You need to have the proper guardrails in place. You need to have, um, one model checking another model to make sure there isn't hallucination going on. Um, You need to often have, uh, uh, for example, you know, it, it being the first round of things and then a human checking it in a second round. So, so, so for example, if you're doing, uh, with just regular AI, we, at Amazon, we did a lot of, um, uh, like, uh, cancer identification from MRIs and, and, and, uh, CAT scans. And, you know, the, the, the ML was about 98% accurate, which is tremendously good. It's not a hundred percent accurate though. So you do still want a doctor to look at it, right? So you have the ML filter ahead of time, and then it goes to the doctor with recommendations. So, um, I, I think there's, and then as the, um, uh, the doctor provides feedback, the model just gets better and better and better. You know, I mean, the reality is this is all just math, right? These, all, all, all ML is just math. It's, you know, Uh, vectors, and it's, uh, tensors, and it's, you know, it's all just math. U…

AI assessment note: “it should be used a lot more, uh, for doing fraud management and financial systems”

Q out there that will sort of get those Waymo's at a tricky situations if I'm not mistaken. And then there's Tesla, which is, um, which is, I would say advancing, but not quite there yet. We don't have autopilot now. So how far away are we? I mean, this is sort of the essential question for autonomous driving conversations. How far away are we from seeing this stuff be mainstreamed?

A That that's, that's, um, so I have two Teslas and I, I play with a full self-driving all the time. It's, uh, it's entertaining, but I, I wouldn't trust it. Entirely, right? If you trust it, you're going to be in trouble. So it's, it's not, it's not a hundred percent there yet. It's a hard problem. It's interesting that you mentioned that the picture on the whiteboard behind me is for a software-defined vehicle and all the different components of the vehicle running across hundreds of thousands of synthetic simulations. And so, um, we work really closely, for example, with NVIDIA on Omniverse. So Omniverse is a synthetic simulator or environmental simulator that Has, uh, full physics and, and full, um, uh, fidelity, and that's really amazing. A lot of the autonomous driving training that has been done in robotics training has been done using Unity and Unreal over time, and those are great environments as well. They look very much like video games when you run them, but people don't watch them. They're all running in the machine memory. Um, and, uh, uh, Omniverse is sort of the first to, To, to go that next level of not being constrained on something that might have to run on a console. So it's, it's pretty amazing. Rev out there. I've been working with him for years on this.

AI assessment note: “It's, it's not, it's not a hundred percent there yet. It's a hard problem.”