On Monday, a team of engineers and executives from Databricks, a data science and AI company, held a virtual meeting via Zoom to find out if their efforts in creating an exceptional artificial intelligence language model had paid off. Their project, named DBRX, was developed over several months and cost approximately $10 million. Similar in design to OpenAI's ChatGPT, DBRX's true capabilities would only be revealed after the final tests. Finally, Jonathan Frankle, the chief neural network architect at Databricks and leader of the DBRX team, shared the exciting news with his colleagues. "We've surpassed everything, " he declared. The team responded with cheers, whoops, and applause emojis. Frankle, who usually steers clear of caffeine, indulged in sips of iced latte after a sleepless night spent compiling the results. Databricks intends to release DBRX under an open-source license, allowing others to build upon their work. Frankle presented data that demonstrated DBRX's superiority over other open-source models in various benchmarks, including general knowledge question answering, reading comprehension, logical puzzle solving, and code generation. DBRX outperformed Meta's Llama 2 and Mistral's Mixtral, two of the most popular AI models available as open-source. "Yes!" exclaimed Ali Ghodsi, the CEO of Databricks, upon seeing the scores. Frankle added that they had even surpassed Elon Musk's Grok AI model, recently open-sourced by his company, xAI. He jokingly mentioned that receiving a mean tweet from Musk would be considered a success. To everyone's surprise, DBRX also closely rivaled GPT-4, OpenAI's closed model that powers ChatGPT and is widely regarded as the pinnacle of machine intelligence.

Frankle beamed, stating that they had set a new state-of-the-art standard for open-source large language models (LLMs). By open-sourcing DBRX, Databricks joins the movement challenging the secretive approach of prominent companies in the generative AI field. OpenAI and Google closely guard the code for their GPT-4 and Gemini large language models, while rivals like Meta have chosen to release their models, believing it will promote innovation and accessibility to more researchers, entrepreneurs, startups, and established businesses. Databricks not only aims to share their open-source model but also intends to disclose the process of creating it. Meta has not provided all the details about the creation of its Llama 2 model. Databricks plans to release a blog post detailing their work and even invited WIRED to observe their engineers making crucial decisions during the multimillion-dollar training process of DBRX. This glimpse highlights the complexity of building leading AI models, but also the potential for reducing costs due to recent advancements in the field. Combined with the availability of open-source models like DBRX, this suggests that AI development will continue at a rapid pace. Ali Farhadi, the CEO of the Allen Institute for AI, emphasized the importance of greater transparency in building and training AI models. As companies increasingly seek an advantage over competitors, the field has become more secretive. Farhadi believes that openness is crucial, particularly when there are concerns about the risks associated with advanced AI models. He expressed his appreciation for any efforts towards transparency and further stated that he expects a significant portion of the market to embrace open models. "We need more of this, " he concluded.