1. Databricks spent $10 million training a generative AI model called DBRX, which is optimized for English language usage but can be customized for other languages.
2. DBRX requires specific hardware to run efficiently, making it challenging for developers and solopreneurs to use unless they are Databricks customers.
3. Despite claims from Databricks that DBRX outperforms other models in certain areas, it falls short of leading generative AI models like OpenAI’s GPT-4 and may have limitations in terms of accuracy and bias.
Databricks has introduced a new generative AI model called DBRX, similar to OpenAI’s GPT series and Google’s Gemini. It was trained over two months and cost roughly $10 million. The model is optimized for English but can translate and converse in other languages. While Databricks claims DBRX outperforms existing open-source models, it is challenging to use unless you are a Databricks customer due to hefty hardware requirements.
DBRX runs up to 2x faster than other models like Llama 2, utilizing a mixture of experts architecture. However, it falls short of OpenAI’s GPT-4 in many areas. Despite efforts to ensure safety and minimize biases, generative AI models like DBRX can still produce inaccurate responses. DBRX is not multimodal and can only generate text, not images.
It is unclear how DBRX was trained and whether copyright or biases are present in the data. Databricks does not currently offer legal fee coverage for IP infringement related to model outputs. The model may struggle to compete with other generative AI models on the market, particularly as it requires expensive hardware and is aimed primarily at Databricks customers. As Databricks continues to refine DBRX, it aims to address challenges related to reliability, safety, and bias in future versions.