Paul Gigot in Conversation with SandboxAQ CEO Jack Hidary

AI and quantum is not your sweet spot, Paul.

We're going to make it all clear today. That's our goal.

Yeah, great question. So we think about the applications of AI, we've got really two buckets. AI for the digital world. That's the LLMs. And so great applications there, documents and images and all the stuff we're doing in the digital world. And journalists and others are making great, great use of LLMs. But you think about AI for the real world. You think about 85% of the GDP of the US and the OECD. You're talking about biopharma making new drugs. We're talking about new materials, new battery chemistries, energy. We're talking about logistics, risk management in banks and funds and others. That's, we need a different kind of AI. Do we want an AI that was trained on cat pictures and Instagram to be doing risk management for a trillion dollar fund? I don't think so. So, we need a different kind of AI. And that's where quantitative models come in. Not language models, but models built and trained on numbers and equations. And then that stay in that pristine lane of focusing on the quantitative part of our world, the molecules, the atoms, the risk management. That's what a large quantitative model is. An LQM, very complementary to LLMs. We need both. And I think soon, one of my predictions I'll share with you on stage here is that I think we're going to see the convergence of the LLMs and LQM in the next 18 months. I think that's coming very, very soon.

Yeah. So great question. We do not download anything from the internet. So that's the first thing and that's good news because we're not downloading copyrighted information from a number of the publications such as has happened in the LLM world. But we really can't download any of that because that is not relevant to the task at hand. If the task at hand is to create a new therapeutic for brain cancer, for glioblastoma, one of the areas we're applying LQMs to right now, then what kind of data do we need? We need two types of data. We need data that come from experiments that are really done in the world. And then we need equation-based data. Sometimes we can call that synthetic data. Data that we create but constrained by the actual equations that govern how electrons and atoms interact with each other. So experimental data, data from real world lab experiments and then synthetic data from the equations themselves. That's the two inputs. In fact, in most of the LQMs we've created, and we now have a few dozen that we've put out, these frontier models we put out, we have created all the data either by contracting with labs to create that data and/or the synthetic data with equations. So we build the data set first. That's the first step. Then we train a model and those models, Paul, are not the models that are used in the language world. The models in the language world are all based on this idea of a transformer. A transformer is a particular architecture of a neural network inspired by all of our brains. If we open up our brains, I started my career as a neuroscientist and I did open up many brains to see how they're architected. And when you look at them, there's about 86, 87 billion neurons in each of our brains. About a hundred trillion to maybe a thousand trillion connections. And this distributed sense of neurons gives us the inspiration to build what we know today as AI. But the specific wiring and architecture of that could be vastly different. And even in our own brains, we have different parts of our brain that process our visual information, that process our auditory information. Those are wired differently. Same here. The LLMs are wired in one way and the LQMs are wired in a different way. So we don't use the transformer models that Google came up with initially for the LLM world. We use a different set of architectures that now become these LQMs and so the data is from a very different source.

Yeah, it's a great question. What we do is we go to first principles. This is first principles work where we're usually working with a domain expert partner. So we ourselves are a technology company. We're not a biotech. We're not a material science company. We're a tech company. And what we do is someone approaches us with domain expertise. In this case, there's a biotech out there with deep domain expertise in glioblastoma. They have that expertise. They have a lot of hypotheses all around what could be a potential drug. We take that hypothesis, that initial molecule and we work it using the equations, using our models to actually optimize it from a billion possible combinations which you'd never want to synthesize in the real world. We try to narrow it down to three to five that we can then get into a lab, synthesize and then test. We've done that with UCSF. We've done that with Sanofi. We've done that with major companies around the world in the bio space. And now we're taking that to the battery space, the chemical space, and now the financial space.

That's right.

We charge them. And the main way we charge is in value-based pricing. The old way of thinking about technology is to say, 'Hey, how long are you using it? How many seats?' That's all old now. The way we do it now is value-based pricing. If we come up with something that's really strong, a great effective therapy, then we get paid. If we don't, we don't. So, we're willing to bet on ourselves. It's a very different model. So if you can take the average glioblastoma patient and extend life with a treatment for six months, two years, three years, you'll be paid on that basis.

So there are milestones along the way, Paul, even before you get to the first patient because typically in the biopharma world it takes 10 to 15 years to get a drug to the clinical trial and then clinical trials have 85% failure. That's initially the area that we're impacting right now.

Yeah. So an example would be UCSF. Stan Prusiner is a Nobel Prize winner there and based on his Nobel Prize winning work he had some hypotheses around drugs that can hit the neurodegenerative diseases. We all know what's going to happen. Every family unfortunately across the world is or is going to be unfortunately impacted by these neurodegenerative diseases. That's Parkinson's, Alzheimer's, dementia. And so he brought us those hypotheses and instead of taking 10, 20 years to try to get those ready for a clinical trial, we've been able to cut down many years off that and make specific predictions about the structure that would be needed of that molecule. And then we validated that in his $21 million cryo-electron microscope, which is quite something to see, I have to tell you. So we've now been able to validate again and again with Santa Fe, with UCSF, with others in the area of Alzheimer's and Parkinson's, in other key areas such as cancer. Those are the two biggest areas that we're having impact right now on the bio side. But what's also exciting, Paul, is we're now extending this to the material world, the material science world, and that's the world that impacts all of us around us. Batteries to store energy. I think we're going to talk about energy today and the energy and the power we need in the US to stay competitive and to beat out other countries out there in terms of the growth we need both in data centers and in manufacturing. Well, that means novel kinds of battery chemistry. Where does that come from? It has to come from an AI that understands our world, that understands the electrons, the atoms as they're functioning in our world. And that's an LQM.

I think we are. If we look at, we just announced with Nvidia a new frontier model. It's called AQ Affinity. A for AI, Q for quantum, affinity for the binding affinity of those drugs, Paul, with the receptors, the targets in the body. And AQ Affinity is now many times faster, hundreds of times faster than anything that the industry has had before in figuring out both the folding, and we're all familiar now with the kind of folding of these proteins, but also the affinity, the amount that it'll bind between the potential drug and the receptor, the target in the body. This AQ Affinity model is one of a number of frontier models we've launched. I do think it's fair to say that this will have significant impact in that time frame of 3 to 5 years.

Well, I think we're all very concerned. Let's start with energy because I think that's really the nub of it. We can build all the data centers we want to, but what's going to power these data centers? I do think we need an all-the-above strategy in terms of the sources of generation of the energy that we need and we also need energy storage. Our grid today, less than 2%, Paul, can be stored in any kind of scalable way. Now that's a problem. If I told you 40 years ago we're going to create something called the web, the internet, and we're going to be able to do computing, you could see the answer, Paul, but you'd have to memorize it because there's no hard drives, there's nowhere to store any data on the internet. You'd say that is the most ludicrous internet I've ever heard of. That's exactly what we have in the energy grid today. We make electrons 24/7, but we cannot store more than 2% of those. And what happens is at night, many of those electrons go to waste. They literally get grounded because otherwise they'll fry the wires that we have going above us. And so we've got to do better in terms of storing so that we can shift the load, load shift the power when we need it, where we need it.

China. Mhm.

That's right.

Large companies come to us and say we need batteries that do not use lithium and it's possible to build these. We need batteries without cobalt for the cathode or anode. Yeah.

I think this is quite urgent if you look at global competitiveness and the need to grow the energy grid and also to make it more resilient. It's one thing to just add more power plants, but the question is can we tap that energy when we need it? If you look at natural gas as an example, natural gas has been growing and that's a good thing in terms of part of the power grid. It's gone up to 40 plus percent now from less than 5% just about 20, 30 years ago. But here's the problem. Right now the backlog to order a turbine to make a new gas plant is what? It's four, five years now. And so that's a problem. We have these key choke points, Paul, in our economy and particularly in our power grid. And we've got to get beyond those.

Yeah. So, what is risk? When we say the word risk, we're talking about a quantification of exposure to an unknown outcome. And traditionally, we've had some tools in statistics to try to get our hands around that. We've had Monte Carlo, we've had some other areas to try to do that. But we now live in a much more complicated world, a nonlinear world, Paul, where the effects of one industry are now impacting the other. And we're seeing that right now, of course, in the oil markets. Oil markets don't just affect the folks in the oil industry, they affect almost every part of the S&P 500. And we think about Procter & Gamble, 75% of the input into Procter & Gamble's products is from the petrochemical supply chain. And so we need models that understand this nonlinearity and that means we've got to use better, more sophisticated quantitative tools when we think about the kind of risk inherent in, say, you have a $200 billion fund, a trillion dollar fund. There's many, many multi-trillion dollar funds today. Schwab, Fidelity, BlackRock, so on and so forth. They could also be creating structured instruments. And so a $1 trillion fund could have a footprint of two or three trillion very easily. And we know what's happened in the past with some of those failures. So the question is, how can we model our economy in a better way? Now, what's the right tool to do that? Is it a large language model trained on videos of people jumping off trampolines? Probably not.