LinkedIn bypassed prompt engineering and instead used model distillation to create its next-generation AI recommender systems, according to Erran Berger, VP of product engineering at LinkedIn, in a recent Beyond the Pilot podcast. The company found that prompt engineering was not a viable option for achieving the necessary accuracy, latency, and efficiency improvements for its job-seeker recommendations.
Instead, LinkedIn developed a detailed product policy document to fine-tune a 7-billion-parameter model, which was then distilled into smaller teacher and student models with hundreds of millions of parameters. This multi-teacher distillation process proved to be a breakthrough, creating a repeatable method now used across LinkedIn's AI products. "There was just no way we were gonna be able to do that through prompting," Berger said. "We didn't even try that for next-gen recommender systems because we realized it was a non-starter."
LinkedIn has been developing AI recommender systems for over 15 years, making it a leader in the field. The company's decision to move beyond off-the-shelf models reflects a growing trend in AI development, where organizations are increasingly tailoring models to specific needs and datasets. Model distillation, the technique employed by LinkedIn, involves training a smaller, more efficient model (the student) to mimic the behavior of a larger, more complex model (the teacher). This approach can significantly reduce computational costs and improve performance in resource-constrained environments.
The implications of LinkedIn's approach extend beyond the realm of job recommendations. The company's success with model distillation demonstrates the potential for organizations to create highly customized AI solutions without relying solely on large, pre-trained models or extensive prompt engineering. This could lead to more accessible and efficient AI applications across various industries.
Berger anticipates significant improvements in the quality of LinkedIn's AI products as a result of adopting this end-to-end evaluation process. "Adopting this eval process end to end will drive substantial quality improvement of the likes we probably haven't seen in years here at LinkedIn," he stated. The company plans to continue refining its model distillation techniques and applying them to other AI-powered features on the platform.
Discussion
Join the conversation
Be the first to comment