Chaque semaine, nous sélectionnons des news Tech, Data et IA pertinentes que nous analysons grâce à l'intelligence artificielle et que nous transformons en podcast (10 minutes) dont les speakers sont des voix générées aussi par IA.

Venez écouter les actus Tech de la semaine ou faire votre veille grâce à ce mini podcast. Inscrivez-vous ici pour ne manquer aucun des prochains podcasts.

😸 DreamGaussian: Generative Gaussian Splatting for Efficient 3D Content Creation

☁️ Oracle Announces Next-Generation Fusion Data Intelligence Platform

🪢 Efficient Streaming Language Models with Attention Sinks

🚀 Digma 1.0 is launching

📱 A Silicon Valley Supergroup Is Coming Together to Create an A.I. Device

‍

Les précédents épisode du podcast Tech Press Review est toujours disponible.

DreamGaussian: Generative Gaussian Splatting for Efficient 3D Content Creation

In a recent article, a new 3D content generation framework called DreamGaussian has been proposed. DreamGaussian aims to combine efficiency and quality in the creation of 3D content. By utilizing a generative 3D Gaussian Splatting model, accompanied by mesh extraction and texture refinement in UV space, DreamGaussian achieves faster convergence for 3D generative tasks compared to existing methods.

To enhance texture quality and facilitate downstream applications, DreamGaussian introduces an algorithm to convert 3D Gaussians into textured meshes and applies a fine-tuning stage for refining details. Extensive experiments demonstrate the efficiency and competitive generation quality of this approach. Notably, DreamGaussian can produce high-quality textured meshes in just 2 minutes from a single-view image, which is approximately 10 times faster than existing methods.

DreamGaussian supports various types of content generation, including image-to-3D and text-to-3D. Furthermore, it has been observed that the method can handle images with a non-zero elevation angle. The optimization progress of DreamGaussian consists of two stages: generative Gaussian Splatting and mesh texture refinement.

The article also presents examples of exported meshes and mesh animations, demonstrating the capabilities of DreamGaussian. These results were achieved using an NVIDIA 3070 (8GB) graphics card.

Overall, DreamGaussian offers a promising solution for efficient and high-quality 3D content creation. Its ability to accelerate the generation process and produce impressive results opens up new possibilities in various domains, such as animation and virtual reality.

Source 👉 https://dreamgaussian.github.io/

‍

Oracle Announces Next-Generation Fusion Data Intelligence Platform

‍

Oracle has introduced the Fusion Data Intelligence Platform, a comprehensive data, analytics, and AI platform designed to help its Fusion Cloud Applications customers improve their business outcomes. By combining data-driven insights with intelligent decisions and actions, the platform aims to provide users with deeper insights and faster time-to-action. The platform includes automated data pipelines, 360-degree data models, interactive analytics, AI/ML models, and intelligent applications. These capabilities run on top of Oracle Cloud Infrastructure (OCI) Data Lakehouse services, offering extensibility at various layers.

The Fusion Data Intelligence Platform addresses common challenges faced by businesses, such as data silos and complex data integration processes. It goes beyond traditional data and analytics applications by providing users with insights that are relevant to their specific roles and workflows. Users can even make decisions and take action directly within the application, without the need to switch between different tools.

Some of the platform's key features include 360-degree data models, prescriptive AI/ML models, rich interactive analytics, and intelligent applications. These features enable organizations to gain a comprehensive understanding of their data and business, automate tasks, explore and visualize data, and make informed decisions. The platform is part of Oracle's long-term vision to help businesses progress from data and analytics to actionable insights. It is not limited to Fusion Cloud Applications and will also be offered for other Oracle industry applications such as health, financial services, and utilities.

Industry analysts have praised the Fusion Data Intelligence Platform for its convergence of analytics, data, and AI. They believe it offers a forward-looking strategy for enterprises seeking to thrive in the data-driven landscape. In addition, the platform includes a range of analytics offerings for different Oracle Fusion Cloud applications, such as ERP, SCM, HCM, and CX, with new additions including accounting, manufacturing, workforce, and customer experience analytics.

Source 👉 https://aithority.com/machine-learning/oracle-announces-next-generation-fusion-data-intelligence-platform/

‍

Efficient Streaming Language Models with Attention Sinks

‍

Researchers have developed an efficient framework called StreamingLLM that allows Large Language Models (LLMs) to handle infinite-length inputs without sacrificing performance. This is particularly useful in streaming applications like multi-round dialogue, where long interactions are common. Traditional LLMs face two challenges in streaming scenarios: extensive memory consumption when caching previous tokens' Key and Value states (KV) during decoding, and the inability to generalize to longer texts than their training sequence length. The researchers introduce the concept of attention sink, where retaining the KV of initial tokens significantly improves the performance of window attention, enabling LLMs to generate coherent text. StreamingLLM is trained with a finite length attention window and can process sequences of up to 4 million tokens. It outperforms the sliding window recomputation baseline by up to 22.2 times in terms of speed. By only retaining the most recent tokens and attention sinks, StreamingLLM ensures efficient and stable language modeling without requiring cache resets. However, StreamingLLM does not expand the LLMs' context window or enhance their long-term memory, making it unsuitable for summarizing extensive texts like books. This framework is ideal for streaming applications, such as daily assistants, that need to operate continuously without relying on past data or consuming excessive memory. StreamingLLM can also be integrated with recent context extension methods. The researchers plan to release the code and data related to StreamingLLM, including the core code, perplexity evaluation, a demo of the Streaming Llama Chatbot, and the StreamEval dataset with evaluation code.

Source 👉 https://github.com/mit-han-lab/streaming-llm

‍

Digma 1.0 is launching

‍

In exciting tech news, Digma 1.0 is officially launching today! Developed by Roni Dover, this platform offers new ways for coders to receive and utilize feedback during the coding process.

Recognizing that developers have access to a wealth of valuable information through observability advancements, Dover and his team set out to solve three key challenges. First, they aimed to eliminate the need for developers to exert cognitive effort in analyzing raw data. To achieve this, they built an intelligent engine that constantly analyzes observability data to identify specific issues in the code. This saves developers from wasting time on investigations that often yield no results.

Secondly, Digma integrates the user experience directly into the Integrated Development Environment (IDE). This means that the feedback data is always within a developer's peripheral vision, allowing for a seamless and efficient coding experience. No more bouncing back and forth between separate dashboards or struggling to interpret complex interfaces.

Lastly, Digma focuses on the code itself. Unlike many other Application Performance Management (APM) solutions that deal with broader metrics, Digma's core focus is on the specific code that needs improvement. It places emphasis on classes, events, methods, properties, as well as synchronous and asynchronous flows within the code.

Dover is eager to share Digma with the developer community and welcomes their thoughts and feedback. This innovative platform aims to revolutionize the way developers utilize feedback during the coding process, making it more efficient and accessible for all.

Source 👉 https://roni-dover.medium.com/digma-1-0-is-launching-d0187fba233f

‍

A Silicon Valley Supergroup Is Coming Together to Create an A.I. Device

‍

OpenAI founder Sam Altman and renowned designer Jony Ive, formerly of Apple, are joining forces to develop a new computing device that goes beyond the smartphone and harnesses the power of artificial intelligence (AI). The aim is to create a device with a different form factor, breaking away from the rectangular screen that has dominated the technology landscape for the past decade. While the project is still in its early stages, Altman and Ive have already come up with some initial concepts. They are also seeking funding of up to $1 billion from SoftBank, a Japanese technology investor known for its support of innovative ventures. SoftBank's involvement would provide access to the semiconductor expertise of Arm, a leading chip design company that SoftBank acquired in 2016 and recently made public. This collaboration between Altman and Ive has the potential to bring about a significant leap forward in AI technology, offering users a new and improved way to interact with devices.

Source 👉 https://www.nytimes.com/2023/09/28/technology/openai-apple-silicon-valley-supergroup-create-ai-device.html

‍