If you had $10 million to raise the profile of your major tech company, how would you spend it? On a Super Bowl ad? An F1 sponsorship?
You could choose to invest in training a generative AI model. While not traditional marketing, generative models are attention grabbers that lead potential customers to a company’s core products and services.
Introducing Databricks’ DBRX, a new generative AI model similar to OpenAI’s GPT series and Google’s Gemini. DBRX is now available on GitHub and Hugging Face for both research and commercial use. With base (DBRX Base) and fine-tuned (DBRX Instruct) versions, users can run and customize DBRX on various types of data.
“DBRX is optimized for English but can also handle conversations and translations in other languages like French, Spanish, and German,” said Naveen Rao, VP of generative AI at Databricks.
Databricks describes DBRX as “open source” like Meta’s Llama 2 and Mistral’s models, sparking debates about the true definition of open source in this context.
Costing $10 million and two months of training, DBRX surpasses existing open source models on benchmarks according to Databricks. However, the challenge lies in accessing and using DBRX, especially for users who are not Databricks customers.
Running DBRX requires specific hardware like Nvidia H100 GPUs, making it costly outside of enterprise budgets. Databricks offers a managed solution called Mosaic AI Foundation Model to overcome these hurdles.
Even though DBRX runs faster than Llama 2 thanks to its MoE architecture, it falls short compared to leading models like OpenAI’s GPT-4. DBRX also has limitations such as potential inaccuracies in responses due to flawed associations during training.
While Databricks continues to refine DBRX, it faces challenges in competing with other generative AI models in the market that offer more features and competitive pricing. The company is exploring ways to address concerns like biases in training data and potential legal implications for users.
Rao assures that Databricks will keep improving DBRX and releasing new versions as they explore new paths in generative AI.
Looking ahead, DBRX has a long journey ahead to catch up with its peers in the generative AI landscape.
This story was corrected to note that the model took two months to train, and removed an incorrect reference to Llama 2 in the fourteenth paragraph. We regret the errors.
