Industrial Generative AI for Industrial-Scale Problems.
It doesn’t take long sitting around with the team from Zapata AI to feel like something magic is going on. Zapata originated in the Harvard chemistry lab of Alán Aspuru-Guzik, a pioneer in applying quantum computing to chemical calculations. He had gathered an impressive group of postdoctoral researchers, who came together in 2016. Physicist Peter Johnson was focusing on mathematically understanding quantum entanglement –the way qubits interact with each other– and developing a new technique to correct errors in quantum computers from environmental interference. Yudong Cao, a computer scientist, was focusing on integrating quantum computing and artificial intelligence, developing methods which use quantum computers for machine learning tasks. Jonny Olson, an expert in quantum optics, developed a machine learning method for discovering properties of molecules only quantum systems could learn –a so-called quantum autoencoder. Rounding out the crew was a fourth-year graduate student Jhonathan Romero Fontalvo, a chemist from Barranquilla, Colombia, who has used quantum computing to analyze the interactions of electrons in molecules.
The group worked so well together, that rather than watching them leave for the likes of Google and Intel, which are developing their own quantum computers, Aspuru-Guzik proposed they create their own quantum computing company. Instead of focusing on making the computers, however, they would focus on creating the applications for them to use. Thus was the birth of Zapata AI, a company dedicated to creating generative AI software using quantum techniques.
Generative AI is on the tip of every business leader’s tongue, and it’s not hard to see why. Large language model (LLM) tools like ChatGPT and Google Bard have heralded a new era of productivity rivaling the productivity gains that followed the launch of the internet.With use cases ranging from marketing content generation, more responsive customer service chatbots, and accelerated research, business leaders are looking to capitalize on the new technology.But while general-purpose tools like ChatGPT might be useful for general-purpose tasks, they aren’t as useful for tasks specific to a business or industry. “For a company that wants to use an LLM to fill in regulatory applications or customs forms, a tool like ChatGPT that’s trained on Ozzy Osbourne lyrics and Aesop’s Fables isn’t necessarily going to be very useful out of the box,” says Christopher Savoie, CEO and Co-founder of Zapata AI.
Companies can fine tune these public or open-source models on their proprietary data, but that comes with its own issues. For one, enterprises in regulated industries often can’t send their sensitive data to a third party LLM provider; they need to build and train their models in house. LLMs also require advanced compute resources like GPUs that can be very expensive with intense energy requirements that leave a large carbon footprint.
Zapata AI aims to address these problems with a new class of enterprise software it calls Industrial Generative AI, which leverages custom generative models trained on enterprise’s proprietary data for their business-specific, industrial-scale problems.
Zapata AI doesn’t see its software as limited to language tasks though. “Language tasks are just the tip of the iceberg for industrial generative AI,” says Christopher, “There are countless use cases for generative models using other kinds of data, including numerical data like time-series data. We use an advanced form of the technology behind ChatGPT to intelligently generate realistic data that fills the gaps missing in real data, even going so far as to create virtual sensors where they aren’t physically possible today. We can also generate new solutions to optimization problems or generate new molecules with the properties we want, for example for drug discovery.” Zapata AI has already found early success executing this virtual sensor concept.
How does Zapata AI do it? With techniques borrowed from the field of quantum computing. Indeed, most of Zapata AI’s innovations in generative AI have been in quantum techniques. In contrast to most quantum computing applications, which require thousands if not millions of fault-tolerant qubits (the quantum equivalent of bits) that may not materialize for another decade at least, Zapata AI’s quantum techniques for generative AI could create immediate value for enterprise use cases —today— using classical hardware.
These techniques inspired by quantum systems but running on classical hardware can be used to compress generative models (think the GPT model powering ChatGPT), making them cheaper and more efficient. They can also be more accurate than uncompressed models of the same size and generalize better than classical models. In other words, these quantum-inspired models generate better data. And with this better data, Zapata AI has, for example, found that automobile manufacturing processes can be made more efficient by minimizing idle time between body, paint, and assembly shops. Given the premise that generative AI will be the fastest path to quantum advantage, Zapata AI plans on pursuing these applications of generative AI to various industry tasks.
But Zapata AI isn’t limiting itself to generative AI — it’s also thinking ahead to the quantum future, working with DARPA to benchmark the quantum resources required for industrially relevant applications. Recent research has even indicated the potential for quantum-enhanced generative models to unlock new drugs for previously untreatable conditions. Zapata AI is primed to be a significant player in this breakthrough field and, more specifically, a trailblazer of a highly efficient, quantum-inspired approach to generative AI for industrial-scale problems.