In today’s edition of Data Points, you’ll learn more about:
- The new MCP spec’s updates for security and long task support
- Runway’s new video generation model’s improved physics
- The impact of Character.AI blocking minors from chat
- Antigravity, Google’s agentic IDE that rivals Cursor and Windsurf
But first:
DeepSeek V3.2 models feature integrated reasoning and tool use
DeepSeek released two models built on three technical components: sparse attention for long-context processing, a new reinforcement learning framework, and an agentic task synthesis pipeline covering over 1,800 environments. DeepSeek V3.2 performs comparably to GPT-5, while V3.2-Speciale performs on par with Gemini 3.0 Pro and achieved gold-medal results in the 2025 International Mathematical Olympiad, International Olympiad in Informatics, ICPC World Finals, and Chinese Mathematical Olympiad. Both models support tool use with reasoning integration, though V3.2-Speciale requires higher token usage and currently lacks tool-call support. Both models’ weights are available under an MIT license. (DeepSeek and Hugging Face)
Google’s AI model outperformed traditional hurricane forecasting
Google DeepMind’s new AI-based hurricane model delivered the most accurate forecasts of the 2025 Atlantic hurricane season, beating established physics-based systems like NOAA’s Global Forecast System. The model correctly predicted the path and Category 5 intensity of Hurricane Melissa a week before landfall, while other models disagreed on the storm’s trajectory. Unlike traditional models that use physics equations to calculate atmospheric behavior, DeepMind analyzes historical weather patterns to identify subtle relationships in past data. The AI model excelled at both track and intensity forecasting — a significant advance, since previous AI systems struggled with intensity predictions. The National Hurricane Center referenced DeepMind in many forecast discussions and expects AI to become a standard component of hurricane forecasting, though meteorologists say the black-box nature of AI outputs means these models will complement rather than replace physics-based systems and human forecasters. Google announced an updated version of its WeatherNext model on November 17. (NPR and Google)
Anthropic boosts security in updated Model Context Protocol spec
Anthropic released a major update to the Model Context Protocol specification, adding support for long-running workflows and simplified authorization. The new Tasks feature lets servers track work across multiple states, enabling use cases like healthcare data analysis and code migration that run for hours instead of timing out. The update replaces Dynamic Client Registration with URL-based registration using OAuth metadata documents, and introduces URL Mode Elicitation so users authenticate in their browser without exposing credentials to the MCP client. The specification now supports sampling with tools, allowing servers to run their own agentic loops using the client’s tokens. (Model Context Protocol Blog)
Runway’s Gen 4.5 takes top spot on video model leaderboard
Runway released Gen 4.5, an AI video generation model that ranks first on Artificial Analysis’s Video Arena leaderboard, ahead of Google’s Veo 3 and OpenAI’s Sora 2 Pro. The model generates high-definition videos from text prompts and excels at physics simulation, human motion, camera movements, and cause-and-effect relationships. The model supports multiple generation modes including text-to-video, image-to-video, keyframes, and video-to-video. Runway acknowledged current limitations including causal reasoning issues, object permanence problems, and success bias in generated actions. Runway is rolling out access, and it should be widely available soon. (Runway and CNBC)
Character.AI restricts open-ended chats for users under 18
Character.AI limited users under 18 to structured interactions starting November 24, while maintaining access to other features like video creation. The AI companion platform, which has 20 million monthly users, faces lawsuits including a wrongful death case from Megan Garcia whose 14-year-old son died after interacting with a chatbot. The company states it has clear thresholds for sexually explicit and violent content and immediately stops conversations at the first detection of self-harm, recommending helplines regardless of user age. Common Sense Media rates Character.AI as “unacceptable” for users under 18, and experts say they will continue testing the platform’s guardrails to ensure teens cannot circumvent the new restrictions. (CNBC)
ICYMI: Google’s Antigravity IDE orchestrates multiple coding agents
Google released Antigravity, a development platform that separates hands-on coding from agent orchestration through two interfaces: an Editor View with AI-powered completions and inline commands, and a Manager Surface for deploying multiple agents across workspaces. Agents execute multi-step tasks across the editor, terminal, and browser — writing code, running applications, and testing features — then surface results as Artifacts like screenshots and task lists instead of execution logs. The platform stores reusable context and code snippets in a knowledge base for future tasks. Antigravity supports Gemini 3 Pro, Claude Sonnet 4.5, and GPT-OSS, and is available free for individuals on MacOS, Windows, and Linux. (Google)
A special offer for our community
DeepLearning.AI recently launched the first-ever subscription plan for our entire course catalog! As a Pro Member, you’ll immediately enjoy access to:
- Over 150 AI courses and specializations from Andrew Ng and industry experts
- Labs and quizzes to test your knowledge
- Projects to share with employers
- Certificates to testify to your new skills
- A community to help you advance at the speed of AI
Enroll now to lock in a year of full access for $25 per month paid upfront, or opt for month-to-month payments at just $30 per month. Both payment options begin with a one week free trial. Explore Pro’s benefits and start building today!
Want to know more about what matters in AI right now?
Read the latest issue of The Batch for in-depth analysis of news and research.
Last week, Andrew Ng talked about the potential AI bubble, highlighting underinvestment in AI applications, the need for more AI infrastructure for inference, and the risks associated with AI infrastructure for model training.
“Despite AI’s low penetration today, infrastructure providers are already struggling to fulfill demand for processing power to generate tokens. Several of my teams are worried about whether we can get enough inference capacity, and both cost and inference throughput are limiting our ability to use even more.”
Read Andrew’s letter here.
Other top AI news and research stories we covered in depth:
- Google led arena leaderboards with Gemini 3 Pro and Nano Banana Pro, showcasing best-in-class multimodal reasoning and image generation capabilities.
- Microsoft and Anthropic formed an alliance, making Claude the first leading language model available from all three cloud giants.
- Record labels backed AI-music startup Klay Image, which secured deals with industry giants Sony, Warner, and Universal.
- Researchers introduced Persona Vectors to help model builders identify and edit out sycophancy, hallucinations, and more.