Databricks Unveils $10 Million Generative AI Model DBRX: What Makes It Stand Out?

Databricks, a leading data and AI startup, launched DBRX, a flexible large language model (LLM) that outperforms all open-source models on industry benchmarks.

DBRX democratizes the training and tweaking of unique, high-performing LLMs for organizations globally, removing the need for closed models. DBRX lets enterprises quickly build, train, and deploy customized LLMs, according to the company's media release.

DBRX, developed by Mosaic AI and trained on the NVIDIA DGX Cloud, uses the MegaBlocks open-source project to improve its mixture-of-experts (MoE) architecture. DBRX enables customized and transparent generative AI solutions for organizations with leading performance and up to double the compute efficiency of competing LLMs.

Is DBRX Better Than ChatGPT?

DBRX outperforms open-source LLMs like Llama 2 70B and Mixtral 8x7B in language comprehension, programming, math, and logic benchmarks. DBRX notably outperforms GPT-3.5 on key benchmarks, creating a new open-source model standard.

However, besides database programming language development, DBRX is behind OpenAI's GPT-4 in most aspects, according to a report from TechCrunch.

Naveen Rao, Databricks' VP of generative AI, concedes that DBRX, like other generative AI models, may "hallucinate." Despite safety testing and red teaming, DBRX's replies may be inaccurate owing to its training to correlate words or phrases with ideas.

DBRX is unable to evaluate or produce visuals, unlike recent flagship generative AI models like Gemini, which are multimodal. Rao says Databricks utilized community-recognized free datasets to train DBRX, but the sources are unknown.

Rao merely said that Databricks has taken safeguards and undertaken exercises to resolve model shortcomings when asked about copyright or bias concerns in the training data. Users may face ethical and legal problems if generative AI algorithms repeat biased or copyrighted data.

Databricks does not provide legal expenses for infringement, but Rao says the business is examining the option.

Given these considerations, some experts predict that DBRX may struggle to attract consumers beyond Databricks clients as its rival OpenAI offers intriguing technology at reasonable pricing. Moreover, many generative AI models match the conventional definition of open source better than DBRX.

Nevertheless, industry leaders acknowledge DBRX's disruptive influence on AI innovation. Accenture's Chief AI Officer, Lan Guan, praises advanced open-source models, while AI2's Principal Software Engineer, Dirk Groeneveld, emphasizes industry transparency and cooperation. Block, Nasdaq, Prosus Group, Replit, and Zoom want to use DBRX to improve their AI and expand.

Databricks Expands Beyond Data Solutions

DBRX is free for research and commercial use via GitHub and Hugging Face. The Databricks Platform, AWS, Google Cloud, and Microsoft Azure allow enterprises to effortlessly incorporate DBRX into their processes to maximize generative AI for their particular business demands.

This development on DBRX generative AI follows Databricks' announcement that it acquired Boston-based applied research business Lilac, which develops data comprehension and manipulation tools.

Last week, the California-based AI firm unveiled plans to merge Lilac's personnel and technology into its data intelligence platform, formerly the data lakehouse, according to VentureBeat. This strategic strategy intends to simplify dataset quality enhancement for production-grade LLM applications across domains.

The acquisition solidifies Databricks' expansion beyond data solutions to include all areas of generative AI. Databricks, established by the creators of Apache Spark, also invested in Mistral, a leading generative AI firm, after acquiring Lilac. Mistral, which raised Europe's biggest seed round last year, is a major generative AI player. Databricks' strategic efforts demonstrate its dedication to providing data management and generative AI solutions.