Information feeds and suggestion methods have lengthy relied on deep studying architectures that rating every candidate merchandise independently. As LLMs have matured, they’ve opened up a basically completely different method, the place a system can motive about content material the best way it causes about language. Nonetheless, that energy comes with a recent set of engineering challenges round value, scale, and analysis.
LinkedIn not too long ago rebuilt its information feed to deal with content material suggestion as a sequence modeling downside. The final method is to foretell what a person will need subsequent, very like an LLM predicts the following token in a sentence.
Tim Jurka has labored at LinkedIn for 13 years and is presently a VP of Engineering. On this episode, Tim joins Matt Merrill to debate how LinkedIn re-engineered its feed, how the workforce combines LLMs with conventional indicators, managing inference prices at huge scale, steering content material high quality utilizing pure language insurance policies, and extra.
Sponsorship inquiries:
[email protected]
