Large language models (LLMs) like OpenAI’s GPT-4 are all the rage these days, owing to their unparalleled ability to analyze and generate text. But for organizations looking to leverage LLMs for specific tasks — say, writing ad copy in a brand’s style — their generalist nature can become a liability.

When the instructions get too precise, even the best LLMs struggle with consistency. Fine-tuning, or narrowing an LLM’s scope, is one solution. But it’s often challenging from a technical standpoint, not to mention costly.

Motivated to find an easier way, a team of researchers from DeepMind, Google, Baidu and Meta founded Reka, which emerged from stealth today with $58 million. DST Global Partners and Radical Ventures led the tranche with participation from strategic partner Snowflake Ventures, alongside a cohort of angel investors that included former GitHub CEO Nat Friedman.

San Francisco-based Reka is the brainchild of Dani Yogatama, Cyprien de Masson, Qi Liu Head and Yi Tay. While working on AI systems including DeepMind’s AlphaCode and Bard, they four co-founders say that they realized it was impractical to expect a large LLM to be deployed for all possible use cases.

“We understand the transformative power of AI and would like to bring the benefits of this technology to the world in a responsible way,” Yogatama told TechCrunch in an email interview. “Reka is a research and product company that develops models to benefit humanity, organizations and enterprises.”

Reka’s first commercial product, Yasa, doesn’t quite meet those lofty ambitions. But it exemplifies the startup’s early approach. Going beyond text, Yasa is a multimodal AI “assistant” trained to understand images, videos and tabular data in addition to words and phrases. It can be used to generate ideas and answer basic questions, Yogatama says, as well as derive insights from a company’s internal data.

In this way, Yasa, which is in closed beta, isn’t dissimilar to models like GPT-4, which can also understand text and images. But the twist is that Yasa can be easily personalized to proprietary data and applications.

“Our technology allows enterprises to benefit from progress in LLMs in a way that satisfies their deployment constraints without requiring a team of in-house expert AI engineers,” Yogatama said.

Yasa is just the start. Next, Reka plans to turn its attention to AI that can accept and generate even more types of data and continuously self-improve, staying up to date without the need for retraining.

To that end, only available to select customers for now, Reka also provides a service to adapt LLMs it developed to custom or proprietary company data sets. Customers can run the “distilled” models on their own infrastructure or via Reka’s API, depending on the application and project constraints.

Reka, it should be noted, isn’t the only startup chasing after models better suited for enterprise use cases. Writer lets customers fine-tune LLMs on their own content and style guides. Contextual AI and LlamaIndex, which recently emerged from stealth, are developing tools to allow companies to add their own data to existing LLMs. And Cohere trains LLMs to customers’ specifications.

Not to be outdone, incumbents like OpenAI now offer tools for fine-tuning models and connecting them to the internet and other sources to ensure that they remain up to date.

But Reka’s sales pitch won over one early customer (and investor), Snowflake, which partnered with the startup to let Snowflake customers deploy Yasa from their accounts. Appen, the big data analytics company, also recently announced that it’s working with Reka to build tailored multimodal model-powered apps for the enterprise.

Rob Toews, a partner at Radical Ventures, had this to say when asked why he invested in Reka:

“What makes Reka unique is how it offers every business the power and potential of an LLM without having to put up with many tradeoffs,” Toews said via email. “Reka’s distilled Yasa models keep the data within the enterprise, they’re incredibly efficient in terms of cost and energy and they don’t require costly research teams building models from scratch. If every business will become an ‘AI’ business, Reka’s ambition is to give each of those businesses its own, production-quality foundation model.”

Yogatama says Reka, which currently isn’t generating revenue, will use its funding to date to acquire computing power from Nvidia and build a business team.

Reka emerges from stealth to build custom AI models for the enterprise by Kyle Wiggers originally published on TechCrunch