engblogs

Accelerating fusion science through learned plasma control using deep reinforcement learning

Using deep reinforcement learning to successfully control nuclear fusion plasma in a tokamak and accelerate fusion science.

Exploring MuZero's application in video compression for YouTube to optimize bitrate and improve video experiences.

MuZero's application in optimizing video compression in the VP9 codec through collaboration with YouTube.

An exploration of using language models to detect and prevent harmful behavior, providing insights and techniques for red teaming and mitigating risks in real-world applications.

Finding harmful model behaviors by automatically generating inputs using language models themselves.

This technical blog post discusses the new features and updates in Chrome 99, including CSS cascade layers, a new picker for input elements, preparations for Chrome 100, origin trials, canvas 2D features, improvements in CSS calc(), and various deprecations and removals.

Exploring strategies for tackling formal math olympiad problems effectively

Introducing GPT-NeoX-20B, a collaborative 20 billion parameter model developed with CoreWeave.

This blogpost discusses three papers from Stanford's Alexa Prize team on improving user experience and behavior. The papers cover topics such as understanding and predicting user dissatisfaction, handling offensive users, and increasing user initiative in socialbot conversations.

Updates and new features in the internal pull-based exception monitoring service Periskop in 2022.

Enhancing language model accuracy through instruction alignment