Lin Qiao Bets the Future Is Smaller Models Tuned on Private Data

Most of the world's data is private, locked inside companies, and Lin Qiao, founder and chief executive of Fireworks AI, believes the next frontier of artificial intelligence lies there.

She argues the public internet, which supplies much of the training for general models along with labeled data, is a small share of the world's data. The majority sits inside businesses and applications and will never be shared, she said, because it is proprietary.

She bets each company will activate its own data to make a specialized model smart about its own problems. Open models make that practical, she argues, because a customer owns the underlying settings, called weights, and can change them.

That case rests on a quality threshold she says has already been crossed. Over the past two years, she said, open and closed models got good enough for many kinds of problems, and open ones became much easier to tune.

With only a small amount of a company's unique data, she said, her team can repeatedly adjust a model until it scores better on that company's own test of its job. Often, she said, the tuned model beats a general purpose model on that specific task.

She points to her own firm as an example. The staff, which she describes as cost conscious, uses open models to help with recruiting and candidate sourcing, to collect feedback, to help with some internal finance work, and to help engineers debug code, she said. Her company runs many such agents internally.