February 21, 2024


Large language models (LLMs) like OpenAI’s GPT-4 are very popular today because of their unparalleled ability to analyze and generate text. But their generalist nature can prove to be a burden for organizations looking to utilize an LL.M. for specific tasks, such as writing advertising copy in a brand style.

Even the best LL.M. can struggle to be consistent when the instruction is too precise. Fine-tuning or narrowing the scope of the LLM is one solution. But this is often challenging from a technical standpoint, not to mention costly.

In search of an easier way, a team of researchers from DeepMind, Google, Baidu, and Meta formed deliciousToday, it’s fetching $58 million. DST Global Partners and Radical Ventures led the round, with participation from strategic partner Snowflake Ventures and a group of angel investors including former GitHub CEO Nat Friedman.

San Francisco-based Reka is the brainchild of Dani Yogatama, Cyprien de Masson, Qi Liu Head and Yi Tay. As they developed AI systems including DeepMind’s AlphaCode and Bard, their four co-founders said they realized it was unrealistic to expect to deploy large LLMs for all possible use cases.

“We understand the transformative power of artificial intelligence and want to bring the benefits of this technology to the world in a responsible manner,” Yogi Tamar told TechCrunch in an email interview. “Reka is a research and product company developing models that benefit people, organizations and businesses.”

Reka’s first commercial product, Yasa, doesn’t quite live up to these lofty ambitions. But it exemplifies the startup’s early approach. In addition to 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, as well as gain insights from a company’s internal data, Yoga Tamar said.

In this way, Yasa, which is in beta, is no different from models such as GPT-4, which can also understand text and images. But the difference is that Yasa can be easily personalized based on proprietary data and applications.

“Our technology enables businesses to benefit from LLM advancements to meet their deployment constraints without requiring an in-house team of expert AI engineers,” said Yogatama.

Asha is just the beginning. Next, Reka plans to turn its attention to artificial intelligence, which can accept and generate more types of data and continuously improve itself to stay current without retraining.

To this end, Reka currently offers its services only to select clients, and Reka also offers a service to adapt the LLMs it develops to custom or proprietary company data sets. Customers can run “refined” models on their own infrastructure or through Reka’s API, depending on application and project constraints.

It should be noted that Reka is not the only startup pursuing a model more suited to enterprise use cases. Writer allows clients to fine-tune the LLM to their own content and style guides. The recent creeps Contextual AI and LlamaIndex are developing tools that allow companies to add their own data to existing LLMs. Cohere trains the LL.M. according to the client’s requirements.

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

But Reka’s sales pitch won over an early customer (and investor) Snowflake, which partnered with the startup to let Snowflake customers deploy Yasa from their accounts. Big data analytics company Appen also recently announced that it is working with Reka to build custom multimodal model-driven applications for enterprises.

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

“Reka is unique in that it offers every business the power and potential of an LL.M. without enduring many of the trade-offs,” Toews said via email. “Reka’s distilled Yasa models keep data in-house, they are very cost- and energy-efficient, and they don’t require expensive research teams to build models from scratch. It is about giving each business its own, production-quality base model.”

Reka, which has not yet generated revenue, will use the funding it has received so far to acquire computing power from Nvidia and build a business team, Yogatama said.