Databricks CEO Ali Ghodsi says small AI models earn more money

Ali Ghodsi, who runs the data company Databricks, said customers often get more from small, task-specific AI than from one giant system. He said so on The Logan Bartlett Show, in an episode posted Nov. 22, 2024.

This matters for anyone whose company pays for AI. Running AI costs real money each time someone uses it. A cheaper system that does one job well could mean lower bills and more tools at work.

Ghodsi said many customers do the same task again and again, like reading insurance documents. For them, he said, a huge general model costs too much and does worse on that one job. He described building AI from many parts that work together, so a team can find which part made a mistake and fix it. He said the surprise at Databricks was this.

"in aggregate, we get more revenue from tiny models."

The quote starts at 11:45. Ghodsi said customers take a big model and shrink it to the smallest version that is good enough.

The story making news now comes from NewsBytes. It reports that Databricks is seeing 100% year-on-year growth in India. NewsBytes says more companies are adopting AI and wanting control of their data. The story doesn't say how much comes from small models.

Ghodsi's company sells these tools, so he has a stake in the claim. He also said he doesn't have a crystal ball, and he expects people to stay involved in checking AI's work for many more years.

Watch the 7 minutes where he explains it.8:18