#170 - new Sora rival, OpenAI robotics, understanding GPT4, AGI by 2027?
Jun 9, 2024
auto_awesome
Discussion on Sora's AI video generator rival, OpenAI's robotics research restart, UAE's AI deals with the US, Zoox testing self-driving cars, and AI collaborations between tech giants
OpenAI is restarting its robotics research group, indicating a focus on advancing AI applications in robotics.
Saudi fund invests in China's effort to develop a rival to OpenAI, highlighting the global competition in AI research and development.
AI models are being developed with an emphasis on interpretability, promoting a better understanding and ethical use of AI technology.
Deep dives
OpenAI Explores Sparse Autoencoders for Concept Extraction from GPT-4
OpenAI released a research paper and blog post on Scaling and Evaluating Sparse Auto-Encoders, aiming to extract concepts from GPT-4. They trained a sparse autoencoder to compress GPT-4 outputs to find 16 million features related to human imperfection, questions, and more. The research includes code for training and interactive feature visualizations.
Stable AI Introduces Stable Audio Open for Sound Design
Stable AI introduced Stable Audio Open, an open-source variant of Stable Audio for sound design tasks. The dataset is trained on non-copyrighted audio data and supports shorter audio clips up to 47 seconds. Stable AI aims to provide interpretable and alignment possibilities in sound design with this release.
OpenAI Research on Sparse Autoencoders for Concept Extraction
OpenAI's research explores the use of sparse autoencoders to extract concepts from GPT-4 outputs. By compressing GPT-4 representations into interpretable features, they aim to identify human imperfection, questions, and more. The research offers insights into scaling sparse autoencoders and evaluating their efficacy for concept extraction.
The Challenges of Achieving Interpretable Models
AI models are being developed with an emphasis on interpretability, challenging the notion of complete opacity within the models. Techniques such as visualizations and feature manipulation are allowing for better understanding and control over model behaviors, potentially aiding in ethical use.
Improving Model Alignment and Robustness
A new approach called 'short-circuiting' is introduced to enhance model alignment and robustness. By directly interfacing with the model at runtime, potentially harmful behaviors can be short-circuited, rather than solely relying on training models to avoid negative actions. This technique aims to preserve and enhance positive behaviors while mitigating harmful ones.
Automating Data Curation for Self-Supervised Learning
A clustering-based approach called automatic data curation is proposed to balance data sets used for self-supervised learning. By clustering diverse data to achieve balanced representation of concepts, the approach aims to provide more efficient and uniform distribution of concepts in data sets, particularly useful for training models effectively.