AI researcher Thomas Scialom from Meta discussed Llama 3.1's development and updates for Llama 4 on the Latent Space podcast. He focused on parameter scaling, scaling law, model architecture, data filtering, and the ongoing work on agent technology.

A recent survey detailed how tech giants like Apple use YouTube subtitles to train AI models. Apple's OpenELM model, released in April, is not used for AI functionalities but aims to support research community. The model is separate from Apple Intelligence and uses publicly available data. No plans for a new version.

Microsoft's design tool, Designer, now available as free app for iOS & Android. Integrated into more apps; generative erasure, color highlighting, automatic cropping, & more. New suggestion templates for generating images. Basic image editing tools & upcoming background replacement feature.

Large language models (LLMs) like GPT-3 and Llama-2 face challenges due to high resource requirements. NVIDIA and UT Austin's FLEXTRON framework offers a breakthrough, optimizing model deployment and resource utilization. The framework uses a flexible architecture, achieving dynamic model adjustment and outperforming other models, making it a milestone in AI technology development. FLEXTRON's innovative technologies enhance adaptability, with experts foreseeing broad AI application in various in

A groundbreaking innovation in AI technology, the Live2Diff system achieves real-time video style transfer at 16 frames per second, revolutionizing digital creativity. Its unique unidirectional attention modeling eliminates the need for future frame data, leading the industry frontier in video AI and reshaping entertainment and AR experiences. While offering unprecedented creative tools, ethical considerations urge responsible AI use. Open-sourcing the system drives industry progress, inspiring

The US considers tightening restrictions on China's chip industry, signaling allies to avoid providing tech to China. FDPR rules may be strengthened, impacting companies like Tokyo Electron and ASML. US is pressuring allies to join in restricting China's semiconductor supply chain. Companies are cautious amid uncertain future directions.

OpenAI launches cost-effective GPT-4o mini, balancing functionality and cost. Excels in multitask language understanding benchmark with 82%, surpassing GPT-3.5 Turbo. Features 128K context window at $0.15 per million input tokens. Plans to add video/audio support. Enhanced security with "instruction-level" mechanism. Market response positive, used by Superhuman and Ramp. Aimed at attracting developers and promoting AI technology.

Mistral AI and NVIDIA collaborate on Mistral NeMo 12B, a high-performance language model for enterprises. It offers versatility, efficiency, multilingual capability, improved tokenization, easy deployment, and hardware flexibility. It is designed with enterprise-grade security and support.

OpenAI released new enterprise control features for ChatGPT Enterprise, including compliance and data security extensions. The Enterprise Compliance API enables auditing of shared data. Integration with third-party compliance providers supports various activities. SCIM will manage user access efficiently.

Groq released top-ranking open-source models Llama-3-Groq-70B and 8B, outperforming major tech giants. Partnership with Glaive on ethical AI training using synthetic data sets new standards for AI industry. Open access via API and Hugging Face drives democratization of AI, reshaping industry landscape.

OpenAI collaborates with Broadcom to develop next-gen AI chips, aiming to enhance computing capabilities. Altman explores partnerships with TSMC and plans to establish a new company for infrastructure investment. The project is still in the early stages, with mass production expected post-2026.

Anthropic launches Claude AI assistant on Android, expanding access to its advanced language models. Features include cross-device conversation sync, real-time image analysis, and language translation. The app is free and compatible with all plans.

RunwayML and Haiper have released advanced video generation models. Haiper 1.5 offers enhanced features, upsampling, and plans for image generation. Despite challenges, Haiper aims to expand and innovate in the AI field.

French AI startup Mistral launches innovative AI models: Mamba architecture for faster processing & Codestral Mamba for code productivity, Mathstral for math reasoning. Models surpass competitors, available for free with Apache 2.0 license. Mistral secures $640M Series B funding.

Tencent's "3D Character Dream Factory" feature, powered by "Tencent Yuanbao," allows easy entry into the 3D creation world. With AI integration, users can create personalized 3D characters, share, edit, and even 3D print them. The technology ensures high precision, naturalness, and realism in character modeling, offering a diverse creative experience. Users can access the feature through the Tencent Yuanbao app or WeChat mini-program, with further applications in Tencent Maps' "Exclusive 3D Car

Allies of Trump are drafting an AI executive order to promote military tech development and reduce regulatory constraints, signaling a pro-Silicon Valley AI policy shift. Republican Party aims to repeal Biden's AI executive order for more innovation-friendly environment. Trump's warming relationship with tech industry includes support from AI companies. Policymakers face challenge of balancing innovation and safety in AI development. Regulatory agencies' attitude towards AI will influence future

ByteDance and Kuaishou are competing in AI video model innovation. ByteDance will share its latest AI model progress on July 19, focusing on long videos and high dynamics. Kuaishou has also made advancements in AI video models, with a user response exceeding seven million videos. This competition showcases the rapid progress and fierce competition in Chinese AI technology for video generation, opening up new possibilities for content creation.

Microsoft's AI-based design app, Designer, offers rich creative features to effortlessly bring imaginative ideas to life. Templates and personalized image generation capabilities cater to amateur and professional designers. Seamlessly integrated with Copilot Pro, Designer enhances work efficiency in Word and PowerPoint. Available on iOS, Android, and Windows, Designer invites creative enthusiasts to explore limitless possibilities.

Samsung Electronics introduces "Sketch to Image" technology, using AI to transform sketches into detailed images on smartphones. Cloud processing offers quick image options, including 3D cartoons. Despite concerns about image authenticity, this innovation may revolutionize art creation and challenge traditional photography principles. The impact on information authenticity is significant as AI expands to more devices.

OpenAI introduces a new technology to improve explanations of large language models like GPT-4 through a game involving a "prover" and "verifier". This method enhances answer accuracy and transparency, crucial for AI applications in healthcare, law, and energy sectors. The game-based training improves model performance, paving the way for more trustworthy and secure AI advancements.