Large-scale programming languages (LLMs) such as OpenAI’s GPT-4 are all the rage these days, due to their unparalleled ability to analyze and generate documents. But for institutions that want to advance LLMs to other jobs — say, writing ads in print — their generalist nature can be a mistake.
When instruction is so precise, even the best LLMs struggle with inconsistency. Streamlining, or reducing the number of LLMs, is one solution. But it is often difficult from a technical point of view, not to mention expensive.
Encouraged to find a simple solution, a group of researchers from DeepMind, Google, Baidu and Meta launched Reka, which went private today with $58 million. DST Global Partners and Radical Ventures led the round with participation from partners Snowflake Ventures, along with a group of angel investors that included former GitHub CEO Nat Friedman.
San Francisco-based Reka is the brainchild of Dani Yogatama, Cyprien de Masson d’Autume, Qi Liu Head and Yi Tay. While working on AI systems, including DeepMind’s AlphaCode and Bard, the four co-founders say they realized that it was not possible to expect a large LLM to be used for all possible applications.
“We understand the transformative power of AI and want to bring the best of this technology to the world in a reliable way,” Yogatama told TechCrunch in an email interview. “Reka is a research and development company that creates models that benefit people, organizations and businesses.”
Reka’s first product, the Yasa, does not meet these lofty goals. But it shows a basic starting point. Beyond text, Yasa is a multi-faceted AI “assistant” trained to understand images, videos and tabular information in addition to text and speech. It can be used to generate ideas and answer basic questions, Yogatama says, as well as information from inside companies.
In this way, Yasa, which is in closed beta, is no different from models like GPT-4, which can also understand text and images. But the twist is that Yasa can be customized with its own data.
“Our technology allows businesses to benefit from the advancement of LLMs in a way that meets their deployment constraints without requiring a team of AI experts,” said Yogatama.
Yasa is just the beginning. Next, Reka plans to turn to an AI that can accept and generate more types of information and continuously improve itself, keeping up with the times without the need for retraining.
To this end, available to selected customers at this time, Reka also offers a service to convert the LLMs they have created into corporate or proprietary datasets. Customers can run “distilled” models in their own builds or through Reka’s API, depending on the usage and project complexity.
Reka, it should be noted, is not new to chasing business-friendly models. Author allows customers to edit LLMs in their own essays and catalogs. Contextual AI and LlamaIndex, recently released, are developing tools to allow companies to add their own data to existing LLMs. And Cohere teaches LLMs to what clients want.
Not to be outdone, regulators like OpenAI are now offering tools to update and connect them to the Internet and other resources to ensure they stay up-to-date.
But Reka’s marketing won over one former customer (and investor), Snowflake, who agreed with the startup to allow Snowflake customers to send Yasa from their accounts. Appen, a big data analytics company, also recently announced that it is working with Reka to develop multimodal-powered business applications.
Rob Toews, a partner at Radical Ventures, had this to say when we asked why he invested in Reka:
What makes Reka unique is how it gives any business the power and capabilities of an LLM without having to endure too many tradeoffs. Reka’s soluble Yasa models save a lot of money in the business, are cost-effective and don’t require expensive research teams from scratch. If every business is going to be an “AI” business, Reka’s desire is to give everyone their business, a kind of creative foundation.
Yogatama says Reka, who is currently not making any money, will use his money so far to acquire computing power from Nvidia and form a business group.
#Reka #hidden #resources #create #models #companies
