Meta’s FAIR team has introduced several new AI developments. One of the highlights is the Meta Video Seal, an invisible yet robust watermark for AI-generated videos that withstands editing and compression. Another tool, Meta Motivo, allows for the control of virtual agents with realistic movements trained through reinforcement learning. Additionally, Meta has introduced Flow Matching as an alternative to diffusion models, already used in applications like Meta Movie Gen and Meta Audiobox.
Meta is also presenting advancements in Theory-of-Mind research and a Large Concept Model to separate language ability from thinking. The Dynamic Byte Latent Transformer is designed to improve processing of individual letters and numbers.
OpenAI has made its search function in ChatGPT free for all registered users worldwide. Previously, this feature was only available to paying customers. The company has enhanced and accelerated the search function over the past months, especially on mobile devices. New integrations include map services and voice search, and users can now set ChatGPT as their default search engine in their browsers. OpenAI has demonstrated how the search works for event recommendations, restaurant tips, or vacation planning, utilizing real-time web data and providing multimedia answers with images, videos, and embedded maps.
The European Union (EU) and companies are facing the challenge of implementing the European regulation for AI systems promptly. Many AI applications subject to the AI Act also fall under other regulations, such as the General Data Protection Regulation (GDPR) and the Digital Services Act (DSA). Additional requirements exist for sectors like finance, medicine, and automotive. A study by legal professor Philipp Hacker for the Bertelsmann Foundation highlights that the AI Act, as a horizontal legal framework, is not sufficiently aligned with sectoral regulations and other digital laws. The AI Act’s broad, risk-based approach aims to categorize AI applications by their risk potential and impose strict requirements on potentially dangerous systems. By August 2026, businesses and member states must gradually implement the AI regulation. However, inconsistencies, overlaps, and ambiguities could hinder smooth implementation and lead to legal uncertainties. The study recommends better coordination of existing regulations to avoid duplications and enhance efficiency. National supervisory authorities should also issue guidelines for applying the AI regulation in specific sectoral contexts.
The BBC has complained to Apple about the AI feature in iOS 18, which incorrectly summarized a headline from the British broadcaster. The headline inaccurately stated that the suspect in a murder case involving a US insurance executive had shot himself. The BBC published an iPhone screenshot as evidence. The BBC emphasized the importance of readers trusting the information they receive, including smartphone notifications, and urged Apple to address the issue. The BBC noted that such misinformation is not an isolated incident, citing other reports from social media. Apple still considers its AI functions as beta.
The USA plans new export rules for AI chips, positioning companies like Google and Microsoft as global gatekeepers for access to this technology outside the USA. The plan, expected to be unveiled in December, requires companies to provide the US government with important information regularly and deny Chinese users access to AI chips. In return, these companies can offer AI capabilities worldwide through their cloud services without additional licenses. Exceptions will be made for the USA’s closest allies, granting 19 countries, including the Netherlands, Japan, and Taiwan, unlimited access to AI chips. Countries under nuclear embargo, like Russia, China, Iran, and Venezuela, remain completely excluded. For other countries, the US government plans strict import limits on powerful AI chips from Nvidia and AMD, with companies needing to compete for licenses to import a limited number of chips into each country.
Google has introduced two new AI models, Veo 2 and Imagen 3, aimed at achieving top performance in video and image generation. The video generation model Veo 2 can create 4K videos and understands cinematic instructions like lens types or camera effects. The video length can extend to several minutes, with fewer hallucinations compared to similar models. Veo 2 initially will be available through selected products like VideoFX, YouTube, and the Vertex-AI platform, with plans to expand to YouTube Shorts and other products by 2025.
The new image generation model Imagen 3 offers improved color balance, more vibrant images, and enhanced detail accuracy. It has been optimized for different art styles and is now globally available through Google’s ImageFX tool in over 100 countries and via Google’s Gemini chat.
Google’s new image generator, Whisk, is still in the testing phase. It creates images by using other images as prompts, without the need for words in the initial step. Whisk, which means “to whisk” in English, aims to mix images and perhaps even froth them up. The example images from Google’s Whisk website appear playful and comic-like, lacking photorealism. Whisk is based on Google’s AI model Gemini and the new image generator Imagen 3. Initially, Gemini describes the prompted images in the background, with the texts then processed by Imagen 3. The underlying prompts can be viewed and edited. Google states that the focus is less on pixel-perfect images and more on quickly bringing ideas to life. However, Whisk is currently only available in the USA and can be accessed exclusively through Google Labs, Google’s testing environment.
Telegram has been blocking around a million groups and channels monthly, using AI among other tools. This initiative follows the arrest of Telegram CEO Pavel Durov in France in August over various allegations, including obstructing official requests for content removal. Durov is currently released on bail. Telegram has launched a new website on content moderation, with an interactive graphic showing a sharp increase in blocked groups and channels since the end of August 2024. The number of blocks has remained high, sometimes exceeding 200,000 per day. Durov stated in a Telegram post that his service blocks around a million groups and channels per month, affecting 10 million users. However, the specific workings of the AI tools remain undisclosed.