How DeepSeek used distillation to train its artificial intelligence model, and what it means for companies such as OpenAI
DeepSeek's sudden emergence has put the AI industry's focus on a technique called distillation.
Chinese artificial intelligence lab DeepSeek roiled markets in January, setting off a massive tech and semiconductor selloff after unveiling AI models that it said were cheaper and more efficient than American ones.
But the underlying fears and breakthroughs that sparked the selling go much deeper than one AI startup. Silicon Valley is now reckoning with a technique in AI development called distillation, one that could upend the AI leaderboard.
Distillation is a process of extracting knowledge from a larger AI model to create a smaller one. It can allow a small team with virtually no resources to make an advanced model.
A leading tech company invests years and millions of dollars developing a top-tier model from scratch. Then a smaller team such as DeepSeek swoops in and trains its own, more specialized model by asking the larger "teacher" model questions. The process creates a new model that's nearly as capable as the big company's model but trains more quickly and efficiently.