ICYMI: Zapata’s Generative AI Keynote at Quantum.Tech Boston

Back in April, Quantum.Tech hosted their annual Boston conference, and Zapata co-founders Christopher Savoie and Yudong Cao gave a keynote address on Generative AI in the Enterprise. They were also joined by MIT’s Jeff Grover, who shared the results of a collaboration with Zapata and BMW using generative AI to optimize BMW’s manufacturing process.  

In case you missed it, you can watch the keynote here, or read on for our recap with added context and links to related resources. 

With a focus on the red-hot field of generative AI, the talk was a departure from the usual quantum computing content at the conference. But it was not incompatible: the talk detailed how quantum techniques can enhance generative AI and unlock new use cases for industry. More than that, the talk presaged Zapata’s current focus on delivering Industrial Generative AI solutions for enterprise. 

By now you’ve probably seen the hype around generative AI and tools like ChatGPT. In 2022 alone, venture capitalists invested $2.6 billion in generative AI startups, and for good reason: McKinsey estimates Generative AI will generate $14T in enterprise value by 2028.  

But language-based tools like ChatGPT are just the beginning of what generative AI can do for enterprise. Generative AI can be used to address complex business problems — and in many cases, quantum techniques can enhance these generative AI solutions. 

Large Language Models (LLMs) for Enterprise 

But first, let’s talk about ChatGPT. Even though it wasn’t the focus of the talk, it’s worth discussing the immense value that can be created by large language models (LLMs) like GPT, the model behind ChatGPT.  

LLMs, leveraging massive neural networks trained on terabytes of data, have an impressive ability to generate satisfying text responses to any user prompt. Potential use cases include support for creating marketing content, legal documents and regulatory filings.  

However, publicly available tools like ChatGPT, Google’s Bard, or Meta’s LLaMA are not ideal for business-specific problems on their own. For one, they’re often not trained on data relevant to enterprise use cases. For example, ChatGPT would not be very useful for a pharmaceutical company looking to fill in paperwork for the FDA or some other regulatory body, since it’s not trained on the company’s proprietary data that would be relevant to the paperwork.

The real value from LLMs for enterprises comes from fine-tuning them — or training them on the data relevant to the task at hand. Prompt engineering can also be used, wherein the relevant data is fed to the model as part of the prompt. Either approach, in isolation or together, can make LLMs much more useful for domain- or business-specific problems.

However, it should be noted that both approaches are labor intensive and require dedicated expertise and infrastructure — both for preparing the training data and iterating on the model over time. In recent months, Zapata has been working to help customers customize LLMs for their business-specific problems — more on this to come.

Generative AI for Analytics

In IndyCar racing, a hundredth of a second could mean the difference between standing on the podium—or watching from the pit. That’s why teams like Andretti Autosport use reams of data collected during the race — as much as 1TB per race — to inform their analytics and ultimately inform their race strategy.  

But they can’t collect data on everything. For example, slip angle (the angle between the direction in which a tire is pointing and the direction in which it is actually traveling) can’t be measured during the race, but it has a major impact on key race decisions like when to make a pit stop.  

Using generative AI, we were able to generate real-time synthetic data for the slip angle, inferring it from historical data and its correlation with other variables that we can measure. As Christopher showed in the presentation, this synthetic data was accurate when compared to real data in an experimental validation to a Mean Squared Error (MSE) of .0863 (100% accuracy being MSE of 0). More information on our work with Andretti can be found here.  

There are many applications outside of racing where synthetic data for unmeasurable variables would be valuable: think reactor attributes in biochemical engineering, risk factors in insurance models, predictive maintenance for industrial equipment. The list goes on. But even if data is not impossible but merely difficult or expensive to collect, this synthetic data can be very valuable. 

The challenge, however, with generating synthetic data from limited data is in generalizing — or generating quality data that doesn’t simply replicate the training data. As it turns out, quantum generative models have advantages when it comes to generalizing from limited data. We demonstrated this in research published this year, which you can read about here.  

One area where the ability to generalize from limited data would be useful is in drug discovery. Imagine you wanted to generate new drug candidates with a generative model, but there were only a small number of existing drugs for the condition that could be used to train the model. Here too, research suggests quantum generative models could have an advantage.  

In work published earlier this year with Insilico Medicine, Foxconn, and the University of Toronto, Zapata scientists explored using hybrid quantum-classical generative adversarial networks (GAN) for small molecule discovery. Not only could the quantum-enhanced GANs generate small molecules, but these molecules had more desirable drug-like properties than those generated by purely classical GANs. You can read more about this research here.

Generative AI for Industrial Optimization

Yet another area where quantum techniques can be valuable for generative modeling is in generating new solutions to industrial optimization problems. By training a generative model on the best existing optimization solutions, the model can propose new solutions that were previously unconsidered. We call this approach Generator-Enhanced Optimization, or GEO.  

BMW learned this firsthand in work published earlier this year in collaboration with researchers from Zapata and MIT, using a quantum-inspired generative model to optimize their manufacturing process. One of those MIT researchers was Jeff Grover, who presented the work alongside Christopher and Yudong at the keynote.  

We won’t go too deep into the details here, since we’ve covered how this works in a recent blog post, and we have a detailed case study of our work with BMW and MIT on our website. But in summary, our quantum-inspired generative model proposed new solutions that tied or outperformed the best traditional optimization algorithms in 71% of problem configurations. GEO’s strongest performance came in problem configurations with large spaces of possible solutions, suggesting it would be ideal for more complex optimization problems.

The advantage could be even greater when you consider that the model used was on classical hardware, not quantum. As Christopher and Yudong shared in the keynote, Zapata research has shown how the quantum-inspired generative models used with BMW can be mapped to real quantum circuits, making them forward-compatible with more powerful future quantum hardware. Thus, companies that implement these quantum-inspired models for their industrial optimization problems will be primed to immediately reap the rewards as quantum hardware matures. 

For more on the advantages of quantum generative AI, see our recent white paper, The Near-term Promise of Quantum Generative AI. And if you’d like to start building generative modeling solutions for the use cases discussed here in your enterprise — LLMs, synthetic data, or optimization — get in touch, we’re always looking for new industrial partners to collaborate with.