Join us in Atlanta on April 10th and explore the landscape of security workforce. We will explore the vision, benefits, and use cases of AI for security teams. Request an invite here.
Databricks, a fast-growing enterprise software company, announced today the release of DBRX, a new open source artificial intelligence model that the company claims sets a new standard for open source AI efficiency and performance.
The model, which contains 132 billion parameters, outperforms leading open source alternatives like Llama 2-70B and Mixtral on key benchmarks measuring language understanding, programming ability, and math skills.
While not matching the raw power of OpenAI’s GPT-4, company executives pitched DBRX as a significantly more capable alternative to GPT-3.5 at a small fraction of the cost.
“We’re excited to share DBRX with the world and drive the industry towards more powerful and efficient open source AI,” said Ali Ghodsi, CEO of Databricks, at a press event on Monday. “While foundation models like GPT-4 are great general-purpose tools, Databricks’ business is building custom models for each client that deeply understand their proprietary data. DBRX shows we can deliver on that.”
Innovative ‘mixture-of-experts’ architecture
A key innovation, according to the Databricks researchers behind DBRX, is the model’s “mixture-of-experts” architecture. Unlike competing models which utilize all of their parameters to generate each word, DBRX contains 16 expert sub-models and dynamically selects the four most relevant for each token. This allows high performance with only 36 billion parameters active at any given time, enabling faster and cheaper operation.
The Mosaic team, a research unit acquired by Databricks last year, developed this approach based on its earlier Mega-MoE work. “The Mosaic team has gotten way better over the years to train foundational AI more efficiently,” Ghodsi said. “We can build these really good AI models fast — DBRX took about two months and cost around $10 million.”
Furthering Databricks’ enterprise AI strategy
By open sourcing DBRX, Databricks aims to establish itself as a leader in cutting-edge AI research and drive broader adoption of its novel architecture. However, the release also supports the company’s primary business of building and hosting custom AI models trained on clients’ private datasets.
Many Databricks customers today rely on models like GPT-3.5 from OpenAI and other providers. But hosting sensitive corporate data with a third party raises security and compliance concerns. “Our customers trust us to handle regulated data across international jurisdictions,” said Ghodsi. “They already have their data in Databricks. With DBRX and Mosaic’s custom model capabilities, they can get the benefits of advanced AI while keeping that data safe.”
Staking a claim amid rising competition
The launch comes as Databricks faces increasing competition in its core data and AI platform business. Snowflake, the data warehousing giant, recently launched a native AI service Cortex that duplicates some Databricks functionality. Meanwhile, incumbent cloud providers like Amazon, Microsoft and Google are racing to add generative AI capabilities across their stacks.
But by staking a claim to state-of-the-art open source research with DBRX, Databricks hopes to position itself as an AI leader and attract data science talent. The move also capitalizes on growing resistance to AI models commercialized by big tech companies, which are seen by some as “black boxes.”
Yet the true test of DBRX’s impact will be in its adoption and the value it creates for Databricks’ customers. As enterprises increasingly seek to harness the power of AI while maintaining control over their proprietary data, Databricks is betting that its unique blend of cutting-edge research and enterprise-grade platform will set it apart.
With DBRX, the company has thrown down the gauntlet, challenging both big tech and open source rivals to match its innovation. The AI wars are heating up, and Databricks is making it clear that it intends to be a major player.