Enlarge image
Databricks boss Ali Ghodsi: The new AI model should score points with low costs
Photo: David Paul Morris/Bloomberg/Getty Images
A spring evening in San Francisco. The table in the “Foreign Cinema” restaurant is richly laid. There's Wagyu beef and fish, plus the "very latest, very fastest" open source AI, says Ali Ghodsi, head of the software company Databricks, which invited people to the event. The company has already specialized in the management and evaluation of large amounts of data. Together with the start-up Mosaic, Databricks has developed its own artificial intelligence. Ten million dollars were invested, and the training alone took two months, says Ghodsi.
The result is the language model DBRX, which is available from this Wednesday. It's not exactly a catchy name, but Ghodsi isn't interested in popularity among the masses either. “DBRX is not a toy, but a tool for companies to work more efficiently,” he says.
Not the best results, but low cost
To prove this, the Databricks CEO throws graphics at the wall. According to this, Databricks LLM performs better according to almost all standard benchmarks than the available open source models from the competition. Even ChatGPT from OpenAI takes longer than DBRX for almost all tasks, says Ghosdi. However, this only applies to version 3.5. The popular ChatGPT 4 continues to be ahead of DBRX in terms of speed and performance.
But not when it comes to costs, says Ghodsi. DBRX saves computing power and therefore money. Overall, they are twice as efficient as the competition – “i.e. only half as expensive,” exclaims Ghodsi. After all, DBRX's source code is open source, i.e. freely available. Every company can download it free of charge and adapt it to their needs. Databricks only makes money from services or computing power that companies buy from it.
Inspiring medium-sized businesses for AI
To gain a foothold in the highly competitive market, Databricks has made important connections. Databricks' investors and partners include Nvidia. The manufacturer has so far equipped the new AI industry almost single-handedly with its special chips. But there is also plenty of competition for AI applications. Microsoft is working on turning its billion-dollar investments into sales, while medium-sized suppliers such as SAP are also investing heavily in the AI business. There are also start-ups that – like Databricks – advertise with freely available tools.
They all target the many traditional companies in this country. These have a lot of historical and therefore valuable data that could be used to train an AI. However, so far there are reservations in many places about sharing this data with large US companies. Open source models could be a way out here. The interest from Europe, it is said internally, is enormous. People in San Francisco are convinced that the performance of common AI language models will soon become increasingly similar. Then it's all about the data that the language models are "fed" with. In the future, they would make the difference between good and bad AI.
However, Ghodsi admits, there must be a certain amount of programming know-how in the company. It is impossible to build your own AI language model without prior knowledge. "But it's not rocket science either." Well at least.