Why ‘distillation’ has become the scariest word for AI companies

livemint.com:

Subscribe to enjoy similar stories. Tech giants have spent billions of dollars on the premise that bigger is better in artificial intelligence. DeepSeek’s breakthrough shows smaller can be just as good.

The Chinese company’s leap into the top ranks of AI makers has sparked heated discussions in Silicon Valley around a process DeepSeek used known as distillation, in which a new system learns from an existing one by asking it hundreds of thousands of questions and analyzing the answers. "It’s sort of like if you got a couple of hours to interview Einstein and you walk out being almost as knowledgeable as him in physics," said Ali Ghodsi, chief executive officer of data management company Databricks. The leading AIs from companies like OpenAI and Anthropic essentially teach themselves from the ground up with huge amounts of raw data—a process that typically takes many months and tens of millions of dollars or more.

By drawing on the results of such work, distillation can create a model that is almost as good in a matter of weeks or even days, for substantially less money. OpenAI said Wednesday that it has seen indications DeepSeek distilled from the AI models that power ChatGPT to build its systems. OpenAI’s terms of service forbid using its AI to develop rival products.

DeepSeek didn’t respond to emails seeking comment. Distillation isn’t a new idea, but DeepSeek’s success with it is raising new doubts about the business models of tech giants and startups spending billions to develop the most advanced AI, including Google, OpenAI, Anthropic and Elon Musk’s xAI. Just last week, OpenAI announced a partnership with SoftBank and others to invest $500 billion in AI infrastructure over the next five years.