The Gradient: Perspectives on AI cover image

Sasha Rush: Building Better NLP Systems

The Gradient: Perspectives on AI

NOTE

Exploring Controllability and Attribute Ability in NLP Systems

The discussion delves into the evolution of controllability in NLP systems, highlighting the shift towards achieving control through the use of good prompts, reflection-based models, and self-criticism for iterative improvements. The exploration extends to the aspiration of generating high-quality outputs akin to GPT models that precisely align with a classifier's output value using diffusion models. There is ongoing interest in research on diffusion models to enhance control properties further. Attribute ability, a new focus area, involves understanding the training data's contribution to specific predictions in large language models, aiming to pinpoint the origination of particular phrases or outputs. The speaker acknowledges that despite advancements, achieving this level of control remains a challenge in current state-of-the-art models.

00:00
Transcript
Play full episode

Remember Everything You Learn from Podcasts

Save insights instantly, chat with episodes, and build lasting knowledge - all powered by AI.
App store bannerPlay store banner