Jack Hidary, SandboxAQ | theCUBE + NYSE Wired: Mixture of Experts

No. No. This is like real deal, Google-

We were nurtured inside of Alphabet, and my big thanks to Sergey, to so many people, Astro Teller, Sergey and Ruth, and many others at Alphabet for nurturing us in our early years. But the intention always was to spin out in order to really grow a large global company, and thank God we're on the way.

When Ruth Porat came over and created a lot more discipline in the company as CFO initially at Alphabet, and also they realized, Larry and Sergey realized that, 'Hey, we really want to expand and do more stuff here.' And Waymo is an example. At that time, it was called Chauffeur. That was the code name for it, Chauffeur. I was lucky enough to be right there and see, really drive in some of the super early autonomous cars. They were not as smooth as they are today. And to see Waymo so successful today, John, is amazing. They just hit half a million paid rides a week just a few months ago. And I think they're on track to hit a million a week, I think by end of year.

And AI or Die, John, just because it's literally that is the choice that people have today. It is for your businesses. Either you embrace this stuff or you're not going to exist. And the AI we're talking about here, John, is not just the AI for the digital world. And I'm sure you've had lots of great conversations here at this table about AI, like large language models such as ChatGPT and Gemini and Anthropic. Great models, really a lot of applications in marketing, customer service, legal, for example, is today's news. But where we focus on at Sandbox is AI for the real world. And so when you want to take AI and make a new medicine for cancer, a new treatment for Alzheimer's. When you want a new battery chemistry to store energy, when you want to do new risk modeling, right? We're here in the New York Stock Exchange right now looking at the options floor. And if you want to do those quantitative models, a language model and something trained on cat pictures and videos isn't really your best friend. You want something trained on numbers and equations, and that's where we shine.

Let's talk about AI for the physical world. We're talking about AI for the real world. That is the new big, big territory for AI. So of course, low hanging fruit a number of years ago when you wanted to train an AI model, the transformer model was the words of the internet, right? The images, the words of the internet. That made a lot of sense. And now, we're seeing the fruits of that labor. ChatGPT, Gemini, Anthropic. These are very powerful, very, very useful tools for productivity, for customer service, marketing. But when it comes to the core of the Fortune 500, the core product development, when you want to make that car, you want to make that spaceship, you want to make that battery, you want to make that drug, that takes a new kind of AI. That's AI for the real world.

Well, exactly, John. Very good analogy to talk about the big wave one that we all experience over a number of decades, because each one of those movements within that big wave led to the next one. If you think about it, people listening to this today may remember dial-up modem. Okay? The crazy sounds. 'Beep, bop, ba, ba, ba.' Okay. That was a dial-up modem. Well, when we're in the era of dial-up modems, you really couldn't have YouTube because it wasn't practical to upload and download all those videos. Once broadband and really almost universal broadband came to the US, to Europe, to other places, that was then the birth of YouTube. Google, of course, bought YouTube and then grew it even bigger. And we have the kind of video world that we have today, be it TikTok, be it others as well. So each wave really opens up the opportunity for the next wave. And that's happening now in this like-

Right. So the first one was the web itself. Tim Berners-Lee is a friend and informal advisor to us, along with Rosemary Berners-Lee as well, who's an official advisor for us. And I joke with Tim, I said, 'Tim, congrats for creating the web. You actually ended up creating the largest training set in human history.' And that's really what 30 years of the web gave us. He gave us this massive training set, and that's great in terms of the large language models. And kudos also to Google. 2017, the paper, Attention is All You Need, which was the paper that defined a new way of architecting neural networks inspired by the brain and having to do that. I really first started using neural networks in medicine when I was working in science at the NIH, bringing neural network ideas to neuroimaging, to brain imaging. So it has a long enrich history there. But now, let's look fast-forward. Okay. So with the low hanging fruit is the worldwide web, download it, make your transformer model, make it work. But the next wave now is really new proprietary data sets, data sets that don't exist on the internet. These are special data sets for biochemistry, for catalysts. Viewers may not know this, but almost everything that you're seeing and touching on this set and in your homes or office is watching this now comes from a catalyst. This laptop you have here could not be made except for these compounds, cold catalysts, which drive a reaction forward quickly with low energy and are not used up in that reaction. It turns out it's really hard, John, to make a catalyst, just like it's hard to make a new drug. AI for chemistry that we have pioneered is now doing that in record time. So this is a very exciting moment now where we're moving from just words as the data set to now quantitative data and equations as the data set. And that's opening up a whole new world of AI that we didn't have before.

So let's start with biopharma, because that's something affects all of us. We all want to see new medicines for cancers. We want to see new medicines for neurodegen diseases. Unfortunately, most everyone viewing this, their family will be impacted at some point by a neurodegen disease, could be Alzheimer's, could be dementia. Unfortunately, this is going to become almost a pandemic as the population ages. And so you want to make a medicine. Today, the traditional way of doing that, it takes 10, 15, 20 years, John, to get to that understanding of what that medicine could be. And then you got to go into clinical trials. And before the advent of AI, we're seeing about 85% failure, John, in clinical trials. So what can AI do about that? Well, what we can do is start to make really good simulations first of the chemical world in a computer. And we could do so with the laws of physics itself. And that is something that we always dreamed of. I dreamed of it more than 20 years ago, okay, in science, but now it's happening. Now, we and others have worked on this and now we have computers that have high fidelity between what we imagine in the computer. And then when we synthesize it, what we see in X-ray crystallography, what we see in cryo electron microscopy, we now see high fidelity between these two. So this is an exciting moment where for the first time we can harness in silico to go for the real world.

We have to. Yes, we absolutely do. We don't use any third-party data in terms of downloading data from the internet or things like that. The first thing we have to do is we have to run lots of experiments. We contract labs in different parts of the world to run thousands and thousands of robotic experiments, things with robot hands, and that is something that you can do out there in the world. We own that data, we take that in, and then we add to it synthetic data based on the laws of physics and chemistry itself. That creates a dataset that's unique, that's proprietary to us. And the next step is we can train models on those. And those models are not large language models, because again, we're not dealing with language here. Those are new kind of models. Those are LQMs, large quantitative models, John.

Well, so we have A for AI.

Yes. We have A for AI. We have the Q for quantitative, but also quantum, because the techniques we're using, we're not running them on a quantum computer because those are not ready yet. We're running them on GPUs and other types of accelerated chips, but we have to use quantum equations. We have to use the equations of quantum physics to analyze and understand the relationship between this molecule and that molecule. That's a quantum issue. And the same now, John, we can go beyond biopharma into batteries, into materials, into oil and gas and those sectors as well.

The special GTC in October. Yeah.

Well, here's a proof point, John.

NVIDIA is not only investor, but also one of our partners. And with NVIDIA and other partners, we announced just a few days ago, AQAffinity, right? AQAffinity is a software platform that we opened up to the world and is now, we believe, the most advanced platform for determining both the structure of those proteins that we want to target in the body, but also the binding affinity between a potential drug and that target protein. So it's doing two things at the same time. 5, 10 years ago, the conventional wisdom, John, is that we would never be able to do that on GPUs.

Out of 1000. Yes, exactly. But we pulled it off last week and we're really proud of that and really proud of our partners. OpenFold3, OpenFold is our partner. It's a consortium of top academics, of Bayer, of many other pharma companies. And of course with NVIDIA, our partner as well, we use the DGX Cloud via GCP in this case, and this is what allowed us to have this massive breakthrough. So this builds on work that we've done with NVIDIA on the SAIR data set, on other data sets. You have to build the really good precision data first. You cannot use random data.

Yeah, we made the dataset public. It's open source right now. It's online right now.

Yes, that's right. That's right.

In fact, in fact, I'm happy to share that NVIDIA just announced that their DiffDock, which is one of their key models for looking at molecular development of new pharmaceuticals is now improved because of the SAIR dataset that we worked on. It's on their website now.

And John, it's a great question. In 2016, when we got going inside of Alphabet, the conventional wisdom was that it would be many decades for us to get the kind of compute that we needed to do real world kind of stuff with molecules, with energy, with risk management. And we made a bet that the various chip operations, both inside Alphabet and at NVIDIA and other places would really ramp up on an almost exponential type of-

And that's what happened.

And so that was the bet that we made. And sure enough, around 2022, 2023 is when the crossover point happened, and we started to see we could do real world computation with these kinds of chips.

We can see breakthroughs, John, both in new therapeutics. We're working on cancer. We made some announcements with iOncologi for brain cancer. This is one of the areas that has been one of the most difficult for traditional oncology. I think we're going to see breakthroughs there. We've talked about our work with UCSF and Stan Prusiner, Nobel Prize winner. That's in terms of neurodegen diseases, in Alzheimer's, dementia, Parkinson's. Very, very important there as well. So those are more important, but also let's look at the energy sector. The energy sector, we need now more than ever, because we need to power buildings like this and we need to power data centers and we need novel battery chemistries. We need novel ways of producing, storing, and transmitting energy as well. And that's all material science. And so AI for material science, John, is to me, one of the new uncharted territories.

Yes. Long horizon investment. Yeah.

Yeah, that's a red alert. Once you see a different R&D lab from the core divisions, red flag on the play.

Yeah. For us, the R&D is part and parcel of the product itself. It is not some separate division out there that we say, 'Hey guys, how are you doing over there in the R&D division and what do you have to throw?'

Yeah. And that's what the lessons we all learn from Xerox PARC and things like that. Massive breakthroughs, but never made it over the wall to commercialization. And so I think the more modern way of doing things is to embed it all together. We've got more than 125 PhDs in our company, very technical company, PhDs in physics, chemistry, biology, AI, mathematics, cybersecurity. And then we've got about 100 engineers, coding engineers, other kinds of engineers on top of the PhDs who are working together with the PhDs side by side to translate their insights into scalable code. And so that's really what this game is about. What it's about is taking domain knowledge that's specific to particular areas, combining across disciplines, that's really powerful. And then the next step is to use code to make it scalable and usable to a huge number of clients.

Oh, these are special folks in our company. I'm super proud of what they've been able to accomplish, John. Every time a challenge is put to them by a customer, they overcome that challenge.

The core four words of our culture are deep impact at scale. The ways on why people come to work every single day, what drives people, what drives me is having significant impact in the world, but at scale, right? And that's what's key now. Right now, because of AI, relatively small teams can have global impact. You've seen that with us, you've seen that with many other companies as well. You don't have to have 10,000 people in your company to have very significant impact. So deep impact at scale is number one. Number two, a learning culture. I love being inside of our company, because every day I'm learning so much. I'm learning for the people around us. We hire people who are way more knowledgeable in so many areas than I am. And so I'm learning a lot. And what I hear from everyone else is they're learning a lot too in this beehive of learning. So I think those are two of the hallmarks that we have there. And also bar racing. This is something that we learn from a number of other company cultures. When you're hiring somebody, it's very easy to fall into the trap of hiring somebody that you know or hiring somebody who knows a friend and so on and so forth. And the bar raiser is somebody who comes in-

That's right. Amazon is one of the examples. They come in from not your team, some other team, sometimes it's our legal counsel, sometimes it's somebody else, and they make sure to ask the hard devil's advocate questions on that hiring committee. And actually, we've turned down 45 people just on that.

Yes, yes, yes, yes. We do the same thing here. And so it's a very powerful tool. Interesting stat, John. We're about 250 people or so in our company. If you go to our website, we have about 20 or so open positions, 25 open positions. How many people have applied to our company in 2025, past 12 months? I asked the team that question, I was expecting maybe a few thousand people had applied. The answer is more than 45,000 people apply to SandboxAQ in the last 12 months. And because people I think want to have impact.

Yeah. People want to have that impact, John.

Yeah. As we bring the world, the physical world, the human world and AI together, this is going to yield incredible benefits for humanity and it's going to take us more than a decade from today to really play this out as a whole industry. Let's take self-driving and autonomous cars to begin with. These are really robotics. That's really what this is. It's not really a car, it's robotics. People mistake Tesla as a car company. It's not a car company. It's a robotics company. And Elon now is making even more clear moves to signal that where he's at least temporarily putting a couple of models on hold to really focus on Optimus and focus on that. But ultimately, what the robotaxi is, it is a robot. And that's what Waymo is as well. It's taken 17, 18 years to get to this stage with Waymo. And now you see that really the scale and also the experience. You've had now the pleasure of being in a Waymo.