Oct 13 (Reuters) – After Intel Corp (INTC.O) bought Naveen Rao and Hanlin Tang's startup in 2016, their job was to help the chip giant make money off artificial intelligence, which requires vast computing power and had become lucrative for the chip industry.
But after leaving Intel last year, the pair on Wednesday revealed a startup called MosaicML. It aims to help businesses and researchers carry out artificial intelligence more efficiently – in other words, by using less computing power and fewer chips. It has raised $37 million in funding from DCVC, Lux Capital, Future Ventures, Playground Global and other venture investors.
MosaicML's founders said that artificial intelligence models, which must be "trained" using huge troves of data, have jumped in complexity since 2018. Training a model can now require million of dollars worth of computing, putting it out of reach for all but large companies.
"Businesses can't do this anymore – they can't have access to the latest methods," Rao told Reuters in an interview. "It's not good for our field."
Rao and Tang linked up with Professor Michael Carbin and Jonathan Frankle, a doctoral candidate in his research group, respectively, at the Massachusetts Institute of Technology, who had published work on how to make machine learning more efficient. What the group found is there is no single technology for reducing AI computing costs, but instead a mosaic – hence the company's name – of different approaches emerging among researchers.
MosaicML puts those existing ideas into a more polished form that is directly usable by businesses. It then also offers paid services, such as tools for estimating the tradeoffs between the speed, cost and accuracy of using different kinds of computing hardware for training an AI model.
David Kanter, executive director of MLCommons, an AI industry group that is not involved with MosaicML, said that predicting and controlling computing costs have become a key issue for businesses looking to tap into the latest advances in AI technology, making work like MosaicML's important to watch.
"Removing hassle, making it faster, making it more affordable, all these things are potentially beneficial to the industry," Kanter said.
(This story corrects titles of Michael Carbin and Jonathan Frankle in paragraph 5)
Reporting by Stephen Nellis in San Francisco; Editing by Steve Orlofsky