Trending AI news stories + paper

Today's Top AI news stories and papers

AI news stories

1. OpenAI’s newest model Sora can generate videos — and they look decent

   OpenAI has introduced Sora, a groundbreaking generative AI model capable of creating 1080p videos from text descriptions or still images, featuring complex scenes with multiple characters and dynamic backgrounds. Sora stands out for its ability to produce videos in various styles, including photorealistic and animated, while maintaining coherence and reducing "AI weirdness," although it does face challenges with complex physics and specific details. Despite its potential, OpenAI is cautious about Sora's public release, focusing on research and exploring safeguards against misuse while considering the integration of provenance metadata in future outputs.

2. Google: Introducing our next-generation model “Gemini 1.5”

   Google and Alphabet have introduced Gemini 1.5, an advanced AI model that significantly enhances performance and efficiency, including a breakthrough in long-context understanding, capable of processing up to 1 million tokens. This next-generation model, including Gemini 1.5 Pro, offers comparable quality to the previous 1.0 Ultra model but with less computational demand, showcasing innovations in architecture and the potential for new AI applications. Gemini 1.5 has undergone extensive ethics and safety testing, ensuring responsible deployment, and is now available in a limited preview for developers and enterprise customers, promising new capabilities for building more helpful and sophisticated AI applications.

3. Slack AI is here, letting you catch up on lengthy threads and unread messages

   Slack is introducing a suite of AI features for its enterprise plans, designed to provide summaries of threads and unread messages, making it easier for users to catch up on workplace discussions without having to sift through extensive message histories. These AI capabilities also include the ability to ask specific questions about workplace projects or policies, with answers generated from relevant Slack messages. This roll-out, currently available in US and UK English for Slack Enterprise users, emphasizes privacy by hosting large language models within Slack, ensuring customer data remains isolated and not used for training purposes.

4. Apple Readies AI Tool to Rival Microsoft’s GitHub Copilot

   Apple is in the process of developing a new artificial intelligence (AI) tool aimed at app developers, positioning itself as a competitor to Microsoft's GitHub Copilot. This initiative is part of Apple's broader efforts to integrate more AI capabilities into its offerings. The tool, which has been in development for the past year, is designed to be included in the next major release of Xcode, Apple's premier programming software. Currently, Apple has expanded internal testing and accelerated development of the tool with the goal of making it available to third-party software makers within the year. This development underscores Apple's commitment to advancing AI technology and enhancing productivity tools for developers.

5. A new way to let AI chatbots converse all day without crashing

   Researchers at MIT and other institutions discovered a unique solution to prevent large language models, like ChatGPT, from crashing during long conversations. Their approach, called StreamingLLM, involves retaining the first few data points in the model's memory cache, which acts as a conversation memory, enabling uninterrupted and efficient conversation even after 4 million words. This method significantly outperforms existing ones by maintaining stable performance and speed, potentially revolutionizing AI applications in various fields by allowing for continuous, long-term interaction without the need for constant resets.

6. Anthropic takes steps to prevent election misinformation

   Anthropic is implementing a technology called Prompt Shield to direct users of its GenAI chatbot, Claude, to authoritative voting information when they inquire about political topics, in anticipation of the 2024 U.S. presidential election. This intervention, developed in response to Claude's inability to provide accurate, real-time election information, redirects users to TurboVote for reliable voting details. The move is part of a broader effort by GenAI vendors, including OpenAI, to prevent election misinformation by restricting their tools' use in political campaigning and steering users towards nonpartisan resources.

AI Repos

1. Stability-AI / StableCascade

2. reorproject / reor

3. NVIDIA / trt-llm-rag-windows