Let’s consider Databricks’ latest offering, DBRX, a generative AI model akin to OpenAI’s GPT series and Google’s Gemini.
If you were looking to elevate the visibility of your major technology company and had a budget of $10 million, how would you allocate it? Would you opt for a Super Bowl advertisement or consider sponsoring an F1 team?
Another avenue worth exploring is investing in training a generative AI model. While unconventional in terms of marketing, generative models have proven to be attention-grabbing assets, increasingly serving as conduits to vendors’ core products and services.
For instance, consider Databricks’ latest offering, DBRX, a generative AI model akin to OpenAI’s GPT series and Google’s Gemini. Accessible on GitHub and the AI development platform Hugging Face for both research and commercial purposes, DBRX comes in two versions: the base model (DBRX Base) and the fine-tuned model (DBRX Instruct). These versions are capable of processing and tuning public, customized, or proprietary data.
Naveen Rao, Databricks’ VP of generative AI, highlighted DBRX’s versatility in providing information across a wide array of topics and languages, including English, French, Spanish, and German.
Although Databricks describes DBRX as “open source,” similar to models like Meta’s Llama 2 and Mistral’s offerings, there’s ongoing debate regarding whether these models truly adhere to the open-source ethos.
Databricks reportedly invested approximately $10 million and two months in training DBRX, claiming superior performance compared to existing open-source models based on standard benchmarks.
However, there’s a caveat from a marketing perspective: accessing DBRX is particularly challenging for those outside of Databricks’ customer base.
DBRX Challenges and Accessibility
To utilize DBRX in its standard configuration, users require a server or PC equipped with at least four Nvidia H100 GPUs, each costing thousands of dollars. While this might be feasible for large enterprises, it’s a significant barrier for individual developers and small businesses.
Although running the model on third-party clouds is an option, the hardware requirements remain steep. Additionally, there are certain restrictions for companies with over 700 million active users, akin to Meta’s limitations for Llama 2 users.
To address these challenges, Databricks offers its Mosaic AI Foundation Model as a managed solution, providing not only access to DBRX but also a training stack for fine-tuning the model on custom data.
Rao emphasized Databricks’ commitment to enhancing the platform’s appeal for custom model building, ultimately aiming to attract more users. DBRX, he explained, serves as a showcase for their pre-training and tuning platform, offering customers an accessible entry point into their generative AI tools.
Future Prospects and Competitive Landscape
Databricks asserts that DBRX runs up to 2x faster than Llama 2, attributed in part to its mixture of experts (MoE) architecture. However, despite its strengths, DBRX falls short of OpenAI’s GPT-4 in many areas, outside of specialized tasks like database programming language generation.
Rao acknowledges DBRX’s limitations, including the potential for erroneous responses due to inherent biases in the training data. Despite efforts to mitigate these risks through safety testing and red teaming, concerns persist regarding the model’s reliability and susceptibility to “hallucinating” answers.
Moreover, DBRX lacks multimodal capabilities, limiting its functionality to text processing and generation.
While Databricks pledges to refine DBRX and release updated versions, the model faces stiff competition from rivals like OpenAI, offering comparable technologies at competitive prices.
In conclusion, while DBRX showcases Databricks’ commitment to advancing generative AI, its current limitations and market positioning suggest a challenging path ahead in a competitive landscape dominated by more established players.
See also: Boeing CEO Dave Calhoun To Step Down In Shakeup Amid Safety Crisis
