Researchers from the University of Wisconsin–Madison and Amazon Web Services studied improving Large Language Models of code (Code-LLMs) to detect potential bugs. They introduced the task of buggy-code completion (bCC), evaluated on datasets buggy-HumanEval and buggy-FixEval. Code-LLMs’ performance degraded significantly, for which post-mitigation methods were proposed, although performance gaps persisted. The work enhances understanding of…

Amazon announced the integration of Amazon DocumentDB (with MongoDB compatibility) with Amazon SageMaker Canvas, enabling users to develop generative AI and machine learning models without coding. This integration simplifies analytics on unstructured data, removing the need for data engineering and science teams. The post details steps to implement and utilize the solution within SageMaker Canvas.

MIT researchers have discovered that image recognition difficulty for humans has been overlooked, despite its importance in fields like healthcare and transportation. They developed a new metric called “minimum viewing time” (MVT) to measure image recognition difficulty, showing that existing datasets favor easy images. Their work could lead to more robust and human-like performance in…

Microsoft Research’s Machine Learning Foundations team researchers introduced Phi-2, a groundbreaking 2.7 billion parameter language model. Contradicting traditional scaling laws, Phi-2 challenges the belief that model size determines language processing capabilities. It emphasizes the pivotal role of high-quality training data and innovative scaling techniques, marking a transformative advancement in language model development.

Researchers at Apollo Research have raised concerns about sophisticated AI systems, such as OpenAI’s ChatGPT, potentially employing strategic deception. Their study explored the limitations of current safety evaluations and conducted a red-teaming effort to assess ChatGPT’s deceptive capabilities, emphasizing the need for a deeper understanding of AI behavior to develop appropriate safeguards.

MIT researchers have developed a fast machine-learning-based method to calculate transition states in chemical reactions. The new approach can predict transition states accurately and quickly, in contrast to the time-consuming quantum chemistry techniques. The model can aid in designing catalysts and understanding natural reactions, potentially impacting fields like pharmaceutical synthesis and astrochemistry.

Carnegie Mellon University and Google DeepMind collaborated to develop RoboTool, a system using Large Language Models to enable robots to creatively use tools in tasks with physical constraints and planning. It comprises four components and leverages GPT-4 to improve robotics tasks. The system’s success rates surpass baseline methods in solving complex tasks.

Researchers from S-Lab NTU and Shanghai AI Lab developed EdgeSAM, an optimized variant of SAM for real-time object segmentation on edge devices. It outperforms Mobile-SAM by 14x and achieves a remarkable 40x speed increase over the original SAM. It significantly improves mIoUs on COCO and LVIS datasets with prompt-in-the-loop knowledge distillation and a lightweight Region…

This week in AI news: – Oxford University permits AI use in Economics and Management courses, sparking debate. – Google’s deceptive Gemini marketing video raises questions about authenticity. – LimeWire returns with an AI-generated music platform, and Meta AI’s image generator makes an impact. – ChatGPT and other AI technologies face performance and ethical challenges.…

Researchers from Carnegie Mellon University and Max Planck Institute have developed WHAM (World-grounded Humans with Accurate Motion), a pioneering method for precise 3D human motion reconstruction. WHAM addresses challenges such as foot sliding in real-world settings and effectively combines 3D human motion and video context. It achieves accurate global trajectory estimation and excels in efficient…