🚀 AI News: Baidu Challenges GPT, Theorem Proving LLMs, MotionGPT & More AI Updates - June 28, 2023 Edition
This newsletter brings AI research news that is much more technical than most resources but still digestible and applicable
Baidu's AI model Ernie 3.5 is taking on OpenAI's ChatGPT and making its presence known, as claimed by Baidu. This chatbot, which builds on Baidu's foundational AI model, Ernie, reportedly outperforms GPT-3.5 in comprehensive ability scores and even GPT-4 in Chinese language tests. Could this mark a shift in AI dominance? We'll keep an eye on the progress of this rising star.
The theorem-proving landscape might see a revolution with Large Language Models (LLMs) proving formal theorems using proof assistants. This has been made more accessible with LeanDojo, an open-source Lean playground. Researchers have also developed ReProver, the first LLM-based prover augmented with retrieval for selecting premises from vast math libraries. This advancement may redefine the role of machine learning in theorem proving.
AI continues to blur the lines between virtual and physical reality. The newly introduced MotionGPT uses multimodal control signals to generate human motion. The approach combines both textual and pose prompts, a unified set of instructions that could bring a new wave of practical and versatile motion generation systems.
UC Berkeley and Meta AI researchers have proposed a novel action recognition model. The model, which fuses 3D pose and contextualized appearance over tracklets, operates under a Lagrangian view of human motion. Their method is showing promising results, achieving state-of-the-art performance on the AVA v2.2 dataset.
Researchers from UC Berkeley have developed vLLM, an open-source library that dramatically accelerates HuggingFace Transformers—24 times faster to be precise. The LMSYS is already harnessing this tool, which allows for higher traffic handling while minimizing operational costs and computational resources.
The video game creation process is about to be revolutionized. Unity has unveiled Muse, an AI-powered assistant for video game creation, which promises a streamlined creation process through the use of natural input like text prompts and sketches.
Robotics may be on the brink of a new era with the introduction of RVT by NVIDIA researchers. RVT, a multi-view transformer for 3D manipulation, takes task language descriptions and camera images as inputs, predicting the gripper pose action. It's proving to be a powerhouse, achieving success in simulations and demonstrating its real-world application with impressive results.
That's all for this edition. Until next time, stay tuned to keep abreast of the most exciting developments in the field of AI!