Artificial Intelligence
Former Fugees member Pras Michél alleges that his lawyer used an AI program called EyeLevel to draft a subpar closing argument in his recent conviction for conspiracy to defraud the U.S. government. Michél’s new legal team argues that the AI tool led to ineffective assistance of counsel. EyeLevel claims to be the first AI used…
Graph and geometric deep learning models have been successful in machine learning for drug discovery, specifically in modeling atomistic interactions, 3D/4D situations, activity and property prediction, and molecular production. However, the lack of large labeled datasets has limited progress. Researchers have created multitask datasets, developed the Graphium machine learning package, and demonstrated the benefits of…
MIT engineers have found that deep generative models (DGMs) used in AI can mimic existing designs but struggle to generate innovative solutions to engineering problems. The study showed that when DGMs were designed with engineering objectives in mind, they produced more innovative and higher-performing designs. The researchers concluded that AI models need to go beyond…
Microsoft is introducing its AI assistant called “Microsoft 365 Copilot” which integrates with ChatGPT and will be available in their office software. The AI tool can generate meeting summaries, draft emails, create Word documents, design PowerPoint presentations, and more. There are concerns about privacy and regulations regarding interaction with AI. Microsoft assures that data processed…
China will be participating in the upcoming UK AI Safety Summit at Bletchley Park, despite initial doubts about their involvement due to security concerns. The summit, which will focus on safety, is the first of its kind globally. The US and China’s geopolitical tensions are increasing, with additional limitations on technology exports from the US.…
Large language models (LLMs) face challenges related to prompt brittleness and biases in the input. Google researchers have proposed a new method called Batch Calibration (BC) to address these issues. BC is a zero-shot approach that minimizes additional computational costs and outperforms previous calibration baselines. It offers state-of-the-art performance, making it a practical solution for…
Daron Acemoglu, an economist at MIT, has been awarded the prestigious A.SK Social Science Award from the WZB Berlin Social Science Center. The award recognizes his influential work on the role of institutions in capitalist economies, the balance between states and societies, and the risks of automation. Acemoglu, who has made significant contributions to labor…
A team of researchers has developed a deep learning compiler for neural network training. The compiler includes a sync-free optimizer, compiler caching, and multi-threaded execution, resulting in significant speedups and resource efficiency compared to traditional approaches. The compiler improves training procedures for real-world applications and has the potential to optimize neural network models across various…
Purina US, a subsidiary of Nestle, used artificial intelligence (AI) and machine learning (ML) to automate animal breed detection on the Petfinder platform. By leveraging Amazon Rekognition Custom Labels, AWS Step Functions, and other AWS services, Purina created an ML model that detects the pet breed from uploaded images and auto-populates pet attributes. This solution…
Flash-Decoding is a groundbreaking technique that improves the efficiency of large language models during the decoding process. It addresses the challenges associated with attention operation, making the models up to 8 times faster. By optimizing GPU utilization, Flash-Decoding reduces operational costs and promotes greater accessibility of these models in various applications. This innovation is a…
21-year-old Luke Farritor, a computer science student at the University of Nebraska-Lincoln, has made a groundbreaking discovery by using a machine-learning algorithm to read the first-ever text from a burnt scroll found in the ancient city of Herculaneum. His breakthrough could lead to the deciphering of numerous currently unreadable ancient texts. Farritor won $40,000 in…
China’s National Information Security Standardization Technical Committee has released a draft document outlining rules for determining problematic generative AI models. The document provides criteria for banning data sources, demands diversification of training materials, and sets requirements for hiring moderators. It also outlines what constitutes prohibited content and addresses the need for more subtle censorship. While…
Researchers have developed RoboHive, a platform for robot learning, to address the challenges in this field. RoboHive serves as a benchmarking and research tool, offering various learning paradigms and hardware integration. Its key features include a wide range of contexts, teleoperation support, visual diversity, clear metrics, and baseline results. The goal is to bridge the…
Nvidia and Foxconn are joining forces to build “AI factories” that will accelerate the production of autonomous electric vehicles (EVs). Foxconn, known for manufacturing Apple’s iPhone, aims to capture 5% of the EV manufacturing market by 2025. The factories will incorporate cutting-edge manufacturing and AI systems to develop and improve EVs. Nvidia’s technologies, including Drive…
Microsoft Azure AI has developed Idea2Img, a self-refinancing multimodal framework for automated image design and generation. Idea2Img utilizes a large language model (GPT-4V) and a text-to-image model to iterate and refine image creation based on user input. The framework demonstrates improved semantic and visual quality in image generation, outperforming other models in user preference studies.
Zephyr 7B alpha outperforms Llama 2 70B Chat on MT Bench. Simple code lines teach you how to run it efficiently.
Researchers from Nvidia and the University of Illinois at Urbana-Champaign have developed Retro 48B, a larger language model that improves on previous retrieval-augmented models. By pre-training with retrieval on a vast corpus, Retro 48B enhances task performance in question answering. The study demonstrates the potential of larger retrieval-augmented models in natural language understanding.
Path planning, a method used to find the best route from one point to another within a map, is often done through search-based planning techniques like A* search. Recent studies highlight the benefits of data-driven path planning, including more efficient discovery of optimal paths and enabling path planning using raw image inputs. This research introduces…
The text discusses the development of a model called Goal Representations for Instruction Following (GRIF), which allows robots to follow instructions and perform tasks. The model combines language and goal-conditioned training to improve performance. The text also provides details on the training process, alignment through contrastive learning, and the evaluation of the GRIF policy. The…
The text discusses the development of a model called GRIF (Goal Representations for Instruction Following) that combines language and goal-conditioned training to improve robot learning. The model uses contrastive learning to align language instructions and goal images, enabling the robot to understand and carry out tasks specified through either language or images. The GRIF model…