Co-Founder of Zapata gives insight into overcoming challenges in the NISQ ERA

Please give us a little introduction on your current role and what you do?

I co-founded Zapata Computing in 2017 and currently serve as the company’s CEO. My main role today is shaping and overseeing the execution of our business strategy -- which is to deliver value today using quantum algorithms with GPUs, other advanced compute, and eventually quantum hardware when it is available and capable.

What are you most excited about for quantum in 2023? What predictions do you have for the year ahead?

What I’m most excited about isn’t actually quantum hardware at all, but rather quantum algorithms running on classical computers. By that, I’m referring to things like tensor networks (TNs), which are a type of data structure that were originally popularized among quantum physicists as a way to simulate quantum states on classical computers. Unlike our competitors, we offer products today that can reduce costs, expand compute capabilities, and offer bottom-line value to customers on today’s classical hardware — not five years in the future.

There are many reasons to be excited about quantum algorithms. Our data shows these algorithms can be used today to compress the size of massive computational models—models that can cost enterprises hundreds of thousands of dollars to operate and have huge carbon footprints. In the context of generative AI, quantum models also have the potential to generate higher quality results. What’s more, many of these quantum models can be run on quantum hardware in the future with minimal modification, which could lead to more substantial advantages as quantum hardware matures.

What kinds of large computational models can be compressed and what would the impact be?

One example of large models that could be compressed with quantum models is Monte Carlo simulation, which is used routinely in the finance industry for things like stress-testing large banks. The recent bank failures are a clear example of why these simulations matter. Our data shows that quantum models can speed up these simulations—which are now so astronomically expensive they can take up half a financial firm’s entire compute budget—by up to 8,400x. The potential savings and additional value from running more risk simulations with this technology are incalculable. We could be stopping global recessions with better, faster, and more complete simulations.

These quantum models also work for the large language models (LLMs) like GPT-4 and other generative AI models that are all the rage today. Our work is showing that we can compress these large models by orders of magnitude without a meaningful drop in performance. In fact, we recently set a new record by compressing the largest generative model ever — a version of GPT in fact. As generative AI gains more traction, compressing these models efficiently and without degrading performance will speed up their runtimes, save enterprises billions in compute costs, and reduce their carbon footprints.

To give an example, imagine a brand that uses AI chatbots for customer service. We’ve all been frustrated by negative experiences dealing with chatbots in the past. If the brand could run higher quality chatbots at a fraction of the cost and time that it would normally take, we’re talking about significantly improving the experience for the brand’s customers.

You also mentioned quantum models could enhance generative AI. How so?

As generative models, quantum models generalize very well from limited data and are more resistant to overfitting than deep neural networks. They can also infer portions of unknown or corrupted data “out of the box”, without needing any special modification, irrespective of where those gaps in the data may be.

To give an example of what this would mean in practice, imagine a generative model designed to propose new drugs. If, for a given medical condition, there weren’t many existing drugs to train the model with, the training data for the model would be very limited. In this scenario a traditional model might just propose the same drugs used to train the model—it would have poor generalization. In contrast, a quantum-based generative model would be better at learning what makes for a good drug given this limited training dataset, and be better at proposing new drugs.

Quantum models also possess flexible methods for generating structured data, which can incorporate a wide variety of constraints arising in real-world problems, particularly optimization problems.

Expand on that, how can quantum models be used for optimization problems?

Many optimization problems have constraints that must be satisfied exactly for the solution to be valid, also known as equality constraints. Without a native way of handling these constraints, optimization solvers can generate vast amounts of invalid solutions that then need to be filtered for valid solutions, making for expensive and inefficient searches. In fact, equality constraints can exponentially reduce the likelihood of finding valid solutions.

In contrast, if we use a quantum generative model to generate optimization solutions, we can encode those equality constraints in the symmetry of the model, in such a way that it only outputs valid solutions. What’s more, while other optimization techniques suffer when you add constraints, more constraints actually improves the performance of the quantum approach since it allows you to use fewer parameters. In other words, you get better performance with fewer computational resources.

In a real-world example of using quantum models for optimization, we recently worked with BMW to apply this generative optimization approach, which we call Generator-Enhanced Optimization, or GEO for short, to a manufacturing plant scheduling optimization problem. We showed that quantum generative models could tie or outperform the best-in-class traditional optimizers in 71% of problem configurations. In particular, GEO delivered the best solutions in configurations when there was a very large space of possible solutions, which can be challenging for traditional optimizers and are widespread in industry.

Register for Quantum.Tech to hear more talks about challenges in quantum.

It’s notable how none of the things you’re talking about use today’s quantum devices. What do you think the key challenge of working with quantum in the NISQ era is?

The key challenge with the NISQ era is that we have yet to find any practical use cases where these noisy, low-qubit devices can outperform classical computers at any scale, in any problem. The devices have too few qubits with too much noise, and we haven’t yet figured out a reliable way to sufficiently correct for the error this produces. There’s also measurement error, which means you need to make more measurements (which means more compute time) for heuristic algorithms. You can show promising research that points to a future quantum advantage, once we have more qubits and error correction, but we know it won’t be for a long time. We have some hints on when that would be based on our benchmarking work with DARPA and customers like bp.

Frankly, we think its problematic that some overly optimistic software companies are claiming they can deliver value today or in the next few years with NISQ solutions, selling PoC projects for hundreds of thousands of dollars that will almost certainly end in failure to supersede classical solutions. That’s going to make quantum computing a hard sell to anybody but the government and the most risk-tolerant enterprises, especially when you have this boom in AI capabilities that are delivering tangible value today.

That’s why we’ve always been focused on how we can try to deliver value to customers today while setting them up for a future quantum advantage — in addition to identifying when that advantage could arrive. We believe the path to value today is with the quantum models I mentioned earlier. While they run on classical computers today, they lend themselves to being easily run on quantum hardware in the future. What this means is that enterprises can build “quantum-ready applications,” which can deliver value today on classical hardware while being forward compatible with the more powerful, error-corrected quantum devices of the future.

Where does your organization sit within the quantum ecosystem?

In many ways we see ourselves as a point of convergence. That could be the convergence between the present and the future. With our work with quantum-enhanced models, we’re bridging the gap between delivering value to customers today while setting them up to harness real quantum hardware as it matures. We also see ourselves as a point of convergence between quantum hardware and software. Our platform Orquestra integrates the best of both hardware and software from across the quantum and classical HPC ecosystems. It allows customers to experiment and benchmark to find the best configuration for their given application or model.

But it’s also bigger than just quantum. We see ourselves at the convergence of three broader trends in computing: the rise of Big Data, Big Models, and Big Compute. Big Data is nothing new, it has been growing for the last two decades. But Big Models and Big Compute are more recent trends. You see models getting bigger all the time, for example with GPT and other large language models now getting into billions of parameters. On the Big Compute side, you have new high performance and exotic compute options than ever with the emergence of GPUs, TPUs, and of course quantum computers. All these trends are synergistic: the Big Data trains the Big Models, and the Big Models run on the Big Compute. We see ourselves as providing customers with a platform to orchestrate their data, models and compute so they can harness all three to tackle their most complex problems at enterprise scale.

In terms of finding the right talent, how are you going about looking for the best people to work on your product?

Being at the convergence point of so many different technologies means we need to look for a diverse talent pool. This includes AI and machine learning experts, data scientists, engineers with enterprise experience, and of course quantum scientists. A lot of the problems we’re dealing with are hard data science and machine learning problems, so we need really smart people that are keeping up with the latest research and can help us apply it in practice. A lot of our new talent is through internships. Our interns have contributed breakthrough research in quantum machine learning and early fault-tolerant quantum computing, and many have gone on to be hired full-time and continue to be leading contributors.

What are you looking to showcase at our show in Boston this April?

Together with our Chief Technology Officer and Co-founder Yudong Cao, I’ll be hosting a keynote session on April 26th at 8:40 AM ET. We’ll show some results from our work using quantum models to compress large computational models as well as present the results of our work with BMW that I alluded to earlier, along with Jeff Grover, our partner on the project from MIT.

If you come by our booth, we’ll be sharing a virtual reality tour of the Race Analytics Command Center, also known as the RACC, which is the nerve center for our work with Andretti Autosport. It’s a really cool experience, and you’ll be able to hear from the great people at Andretti about the work we’re doing to apply advanced machine learning techniques to give them an edge in their race strategy.

Who are you looking forward to meeting at the conference?

We’re hoping to meet with companies that are looking for solutions that work today—those that don’t want to wait for the era of fault-tolerant quantum computing to see an improvement. People that are willing to use quantum algorithms on classical computers today while the real quantum hardware matures. Particularly organizations that are already using large machine learning models, we can help them reduce the costs of training and running those models today. We want to talk with companies that have large, expensive models and huge problems and are a looking for ways to do solve those problems faster and at a lower cost.

Zapata will be attending Qantum.Tech USA on April 25-26. to see them at the event, register here.

Subscribe to Our Free Newsletter