Are you ready to bring more awareness to your brand? Consider becoming a sponsor for The AI Impact Tour. Learn more about the opportunities here.

In what may be a first, an Eastern European bank is teaming up with the government of Hungary to field an AI supercomputer that will be used to create a large language model of the Hungarian language.

OTP Bank, which was founded in Hungary and operates banks across the region, worked out an agreement under which the government provides about half the funding for the supercomputer developed under contract with SambaNova Systems. The government will have access to the system for public and academic research, said Péter Csányi, deputy CEO and head of the digital division at OTP.

The contract with SambaNova Systems capitalizes on the Generative Pretrained Transformer (GPT) for generating large language models that SambaNova announced in October, branding it as dataflow-as-a-service. “Building a system like this to run a GPT model is not something any bank has done before,” said Marshall Choy, VP of product at SambaNova.

The bank decided that creating AI applications for the Hungarian language was in its own self-interest in an increasingly digital economy, Csányi said. “Our capability to adapt to a changing world is one of the core capabilities that we need to bring in-house.”

VB Event

The AI Impact Tour

Connect with the enterprise AI community at VentureBeat’s AI Impact Tour coming to a city near you!


Learn More

Further, OTP Bank and the government agreed that the Hungarian language was unlikely to get its own AI language model anytime soon if they did not take the initiative because it is a hard language to learn and one that relatively few people speak. The experience of working on this project will also give the bank expertise it can apply to generating language models for the languages of other countries in the region, he said.

The AI supercomputer, which SambaNova says will be the fastest in Europe, is scheduled to go live in 2022.

Nation-sized needs

“We’ve been observing the trends for a couple [of] years now, where things are definitely trending towards larger and larger models, which require larger and larger AI supercomputer-like resources to build. That’s why we’ve productized this,” Choy said. SambaNova’s product allows enterprises to create their own GPT models as an alternative to relying on shared resources like OpenAI’s GPT-3.

“We’re also significantly reducing the burden on Péter and his team to go and hire hundreds of data science and machine learning professionals. In fact, he’s going to be able to do this with a dozen or so folks,” Choy said.

The GPT approach to creating large language models uses AI deep learning techniques to allow the software to discover patterns within a language without being explicitly trained on them. That’s the “generative” part, allowing the software to write its own rules based on analysis of large volumes of text content from that language. While large language models speed up development and reduce the labor required to produce a model, they also have pitfalls, such as a tendency to incorporate biases and falsehoods derived from the text they consume.

Whatever their flaws, a recent State of AI report found large language models to be so significant that many nations are “nationalizing” research into them for fear of being left behind.

Businesses are equally concerned with keeping pace, Choy said. “AI, we believe, is going to have a refactoring effect not limited to banking but all industries, similar to what the internet did 20 years ago.”

Csányi said he was skeptical when members of his IT team first came to him with the idea, thinking the budget and talent required would be beyond the bank’s reach. “What SambaNova has done is make this accessible at a reasonable cost,” he said. And while developing the talent within IT to work with AI models will be a challenge, he thinks the bigger challenge will be preventing his people from devoting all their time to it and neglecting the bank’s routine operational needs.

OTP Bank expects to put natural language processing applications to work for things like customer service, fraud prevention, loan origination, and cybersecurity — as well as purposes that may not become obvious until the technology is in production, Csányi said. “We have no shortage of use cases.”

VentureBeat's mission is to be a digital town square for technical decision-makers to gain knowledge about transformative enterprise technology and transact. Discover our Briefings.