How model customization is bringing generative AI to the enterprise

There is a lot of excitement and activity around generative AI models, especially in the short time since the launch of ChatGPT.

ChatGPT—and the GPT-3 large language model (LLM) on which it is based—were trained on public data, which serves as an excellent foundation for consumer applications, but lacks the personalization, privacy, and security that an enterprise requires. That’s where Rodrigo Liang, co-founder and CEO of SambaNova Systems, is looking to make a difference with today’s launch of his company’s SambaNova Suite, which aims to help companies build and deploy custom models of generative AI.

>>Follow VentureBeat’s ongoing coverage of generative AI<

SambaNova started out in 2017 focused primarily on AI hardware, raising an impressive $676 million in April 2021 to support its efforts. In recent years, the company has expanded beyond its initial focus on hardware to also develop support for machine learning training and inference across different models with its Dataflow-as-a-Service offering. The new SambaNova Suite expands the dataflow service with a collection of capabilities that allow organizations to customize open source and proprietary generative AI models to meet their specific requirements.

“The focus of SambaNova is how to bring more generative AI capabilities to the enterprise,” Liang told VentureBeat. “There are certain things that companies need to be successful, and we are doing it today for them.”

Nvidia is not the only AI hardware vendor for generative AI

There is a growing list of vendors that rely on generative AI models to help enable enterprise use cases.

Content creation is a particularly vibrant field of enterprise generative AI customization. Jasper AI, for example, recently announced its Jasper for Business offering, designed to help customize AI for a specific business. Typeface emerged from stealth on Monday with its enterprise content creation platform for generative AI, along with $65 million in funding. Quantive last week announced its foray into generative AI, helping organizations with their business strategy.

What sets SambaNova apart from others in the generative AI space is that it has its own hardware to help streamline enterprise use cases. Rather than relying solely on Nvidia GPUs like many in the industry, SambaNova has developed its own custom silicon optimized for machine learning training and inference.

“What we did was … take the AI ​​software stack that people really want to use with PyTorch, TensorFlow and complex models like GPT, and strip it down to silicon,” Liang said. “We have silicon that was really custom-built to run these big models for the company versus the other way around.”

The team behind SambaNova, including Liang, have experience building microprocessors for Sun Microsystems and Oracle. Liang said SambaNova created an AI microprocessor that is extremely energy efficient and performant to run these enterprise generative AI applications.

Liang emphasized that custom silicon also allows for continuous training and inference capabilities, so the data that feeds generative AI models can be kept up-to-date.

“In business, you need real-time information, so you don’t want your models to really lag behind,” he said.

Privacy, Responsible AI, and the Enterprise

With the new SambaNova Suite, Liang said his company is looking to solve the key challenges companies face with generative AI. Among these challenges are customizing a company’s specific data, as well as the ability to limit bias and provide accountable and explainable AI.

With its platform, SambaNova allows an organization to run custom training in a private environment on any data the organization has, including unstructured data that can be found in a company’s Slack discussion channels. Taking it a step further, Liang said the platform provides organizations with transparency into how a given AI model actually works.

“SambaNova was built to be able to show exactly how the model arrived at a given conclusion,” said Liang. “We have all the processes stored on how we trained and tuned the model, so that when an auditor comes in or someone wants to check if there is bias or why something happened a certain way, you can actually work through the flow and check if your results were correct. done right.”