Large language model
Apple researchers have introduced Matryoshka Diffusion Models (MDM), a family of diffusion models designed for high-resolution image and video synthesis. MDM utilizes a Nested UNet architecture in a multi-resolution diffusion process to process and produce images with varying levels of detail. The training plan progresses gradually to higher resolutions, demonstrating robust zero-shot generalization and high-quality…
This text discusses the installation and use of Rook Ceph as a replicated storage class for Kubernetes clusters. It provides step-by-step instructions on how to deploy Rook Ceph, create storage classes, deploy a file-sharing app, and test the resiliency of the storage solution. The article concludes by highlighting the scalability and reliability of Rook Ceph…
This article discusses the evolution of Large Language Models (LLMs) for code, from RNNs to Transformers. It covers the development of models like Code2Vec, CodeBERT, Codex, CodeT5, PLBART, and the latest model, Code Llama. These models have advanced code understanding and generation tasks, improving programming efficiency.
Researchers from Tsinghua University and Zhipu.AI have released an open-source bilingual language model called GLM-130B with 130B parameters. GLM-130B outperforms GPT-3 and PaLM on various benchmarks, achieving a zero-shot accuracy of 80.2% on LAMBADA. The researchers also shared their training process and experiences, highlighting their commitment to transparency in language model development.
According to an academic, Artificial Intelligence (AI) and algorithms have the potential to fuel racism, political instability, polarization, and radicalization. These technologies, which are not limited to national security agencies, can contribute to political violence and pose a threat to national security.
The HUB framework, developed by researchers from UC Berkeley and Stanford, addresses the challenge of integrating human feedback into reinforcement learning systems. It introduces a structured approach to teacher selection, actively querying teachers to enhance the accuracy of utility function estimation. The framework has shown promise in real-world domains such as paper recommendations and COVID-19…
A physical neural network has achieved a milestone in machine intelligence by learning and retaining information in a manner similar to human brain neurons. This breakthrough paves the way for the development of efficient and low-energy machine intelligence for complex real-world learning and memory tasks.
Large language models (LLMs) have gained popularity in the AI community as they are seen as a step towards artificial general intelligence (AGI). However, LLMs have limitations, such as dependence on unstructured text and difficulty integrating new knowledge. Researchers are exploring the use of graph-structured data to address these issues. Google Research has conducted investigations…
MATHVISTA is a benchmark to assess the mathematical reasoning abilities of Large Language Models and Large Multimodal Models within visual contexts. It combines various mathematical and graphical tasks and includes existing and new datasets. The benchmark reveals a performance gap compared to humans and emphasizes the need for further advancement in AI agents with mathematical…
This text reviews the current top open-source language models available.
YouTube Music has launched a new feature that allows users to create personalized playlist cover art using generative AI technology. Users can select a theme and specific request, and YouTube’s AI system generates a selection of images to choose from. This feature is currently available to English-language users in the United States but will expand…
Progressive Conditional Diffusion Models (PCDMs) have been introduced by Tencent AI Lab to address the challenges in pose-guided person image synthesis. PCDMs consist of three stages: predicting global features, establishing dense correspondences, and refining images. The method effectively aligns source and target images at multiple levels, producing high-quality and realistic results. It also demonstrates improved…
This article discusses three key questions for junior data scientists to consider when thinking about their future careers. The first question is whether they want to be an individual contributor, a manager, or a combination of both. The second question is whether they want to specialize in areas like machine learning, decision science, or analytics…
Researchers from JPMorgan Chase & Co. conducted an experiment using OpenAI’s GPT-4 model to determine if it could pass the CFA exam. They found that ChatGPT would likely not be able to pass the CFA Levels I and II, while GPT-4 had a decent chance with appropriate prompting. Both models faced challenges with Level II.…
On day two of the AI Safety Summit, UK Prime Minister Rishi Sunak announced that industry leaders such as Meta, Google Deep Mind, and OpenAI have agreed to allow government evaluation of their AI tools before market launch. The summit also established the AI Safety Institute and unveiled a forthcoming “state of AI science” report…
Large language models (LLMs) are being used more frequently as conversational systems, leading to increased reliance on them for answers. To understand how these models respond to questions about ongoing debates, we need datasets with human-annotated labels reflecting contemporary discussions. To address this, we propose a new way of creating a dataset for controversial questions.
Stanford University researchers have introduced EquivAct, a visuomotor policy learning approach that enables robots to generalize tasks across different scales and orientations. The proposed method incorporates equivariance into the visual object representation and policy architecture to ensure robustness across variations in object placements, orientations, and sizes. By using SIM(3)-equivariant network architectures, the learnt policy can…
Researchers from the Universities of Oxford, Münster, Heidelberg, and Exeter have developed innovative photonic-electronic hardware capable of handling three-dimensional (3D) data. This breakthrough significantly enhances the parallelism of data processing for artificial intelligence (AI) tasks. By using radio-frequency modulation, wavelength multiplexing, and non-volatile memories, the team achieved a high level of parallelism, surpassing previous accomplishments.…
Amazon SageMaker Canvas now supports deploying ML models to real-time inferencing endpoints, eliminating the need for manual export, configuration, testing, and deployment. This feature enables users to easily consume model predictions and drive actions outside of the SageMaker Canvas workspace. The process of deploying a model in SageMaker Canvas to a real-time endpoint is explained…
Teachers and students can use a generative AI solution to create course materials and learn English words and sentences. The solution provides real-time assessments and personalized feedback for students. Teachers can generate questions and answers, create images, save assignments to a database, and browse existing assignments. Students can select assignments, answer questions, check grading scores,…