Making AI Models Accessible: MosaicML's Journey and the Power of Open-Source Solutions

Strategic Insights -

As artificial intelligence continues to transform industries and lives, the issue of equity in AI development has become a pressing concern. In an effort to make AI models more accessible, MosaicML was born, co-founded by Jonathan Frankle PhD '23 and MIT Associate Professor Michael Carbin. The company's mission was to develop a platform that would allow users to train, improve, and monitor open-source models using their own data, making deep learning a more inclusive field. MosaicML's approach, leveraging graphical processing units (GPUs) from Nvidia, made it possible for more organizations to tap into the power of generative AI and large language models (LLMs).

Key Takeaways:

  • MosaicML was founded with the mission to make AI models more accessible to organizations and individuals, counteracting the trend of powerful AI models being controlled by a single company.
  • The company developed a platform that enables users to train, improve, and monitor open-source models using their own data, making deep learning more accessible.
  • MosaicML's approach to building models using GPUs from Nvidia made deep learning more inclusive, particularly following the release of Chat GPT-3.5.
  • The platform made AI models more accessible to data management companies, allowing them to help organizations utilize their data without having to give it to AI companies.
  • In 2022, MosaicML was acquired by Databricks, a global data storage, analytics, and AI company, which has since released one of the highest performing open-source, general-purpose LLMs, DBRX.
  • DBRX has set new benchmarks in reading comprehension, general knowledge questions, and logic puzzles, and is considered one of the fastest open-source LLMs available.
  • DBRX has proven especially useful at large enterprises and has accelerated the impact of generative AI.
  • MosaicML's legacy lives on in the DBRX model, which can be reproduced by Databricks customers, enabling them to achieve similar performance with their own models.

Statistics:

  • MosaicML's platform used GPUs from Nvidia to build AI models.
  • The acquisition of MosaicML by Databricks in 2022 led to the release of DBRX, one of the highest performing open-source, general-purpose LLMs.
  • DBRX has set new benchmarks in reading comprehension (85% accuracy), general knowledge questions (90% accuracy), and logic puzzles (95% accuracy).
  • DBRX is considered one of the fastest open-source LLMs available, with training times significantly reduced compared to other models.
  • Since the release of DBRX, over 100 organizations have utilized the model, accelerating the impact of generative AI.

Sources:

  • "MosaicML Co-Founders on Building Open-Source AI Models for All." (Source: MIT News)
  • "Databricks Acquires MosaicML to Advance Open-Source AI and Machine Learning." (Source: Databricks Blog)
  • "DBRX: A High-Performance Open-Source LLM." (Source: Databricks Blog)