AI’s Inflection Point: How MosaicML is making AI models more accessible

By Jéssica Leão and Ambar Bhattacharyya

We’re excited to announce Maverick Ventures’ investment in MosaicML, a company at the forefront of the generative AI space and our ninth investment in an AI-driven innovator.

What does MosaicML do?

MosaicML’s products make compute useful for large-scale ML training, enabling enterprises to access state-of-the-art AI/ML tools, methods, and infrastructure, and putting the power of large models in the customer’s hands.

Think: any enterprise — whether a bank, media company, or SaaS vendor, can integrate a fully functioning AI model into their products, built on their proprietary data, and done so cheaply and efficiently.

MosaicML’s secret sauce is an optimization of systems and algorithms, orchestration for improved resource utilization and resilience, as well as packaging for ease of use, that makes every dollar spent on computing when training large ML models many multiples more efficient and effective. And this isn’t just talk: Mosaic published the costs to train compute-optimal LLMs, all measured on the MosaicML Cloud, and the profiling shows that a compute-optimal model, with the recipe to reach the same quality as the original GPT-3, can be trained for less than $500K.

MosaicML also recently partnered with Stanford to create BioMedLM, a purpose-built AI model trained to interpret biomedical language, and the cost of training this quality model from scratch was only ~$38,000! Mosaic also put itself to the test with MLPerf, and submitted results with their algorithmic stack to both ResNet50 and BERT benchmarks, and won them both (see: ResNet results and BERT results).

The MosaicML Team

The caliber of people building Mosaic is first-class. Naveen, co-founder and CEO of MosaicML, is a second-time founder, formerly the founder and CEO of Nervana Systems, a deep learning ML platform. Nervana designed machine learning chips and software architecture, and was early to the deep learning space by launching a hosted cloud platforms for “deep learning on demand.” Nervana was later acquired by Intel, where Naveen spent close to four years leading the AI vision, most recently as the Corp VP and General Manager of the Artificial Intelligence Products Group.

Building MosaicML alongside Naveen are his co-founders Hanlin Tang, Mosaic’s Chief Technology Officer and former Director of Intel AI Lab; and Jonathan Frankle, Chief Science Officer and author of seminal paper on efficient deep learning: The Lottery Ticket Hypothesis.

We are thrilled to back this exceptional team. Their efforts show: user feedback is glowing. Customers highlight how they’ve experienced astonishing results with MosaicML, which allows them develop models that they thought were impossible to build. And for those who were already on their AI journey, they report that with MosaicML, what would have taken months now happens in a matter of 1–2 days.

Our Partnership

For all these reasons, and many more, we are incredibly proud to announce our investment in MosaicML. We know this will be a partnership for years to come, and we cannot wait to them continue to build towards their vision of making AI models more accessible for everyone.

The views expressed herein are solely the views of the author(s) and are not necessarily the views of Maverick Capital, Ltd. or any of its affiliates. They are not intended to provide, and should not be relied upon for, investment advice.