In today’s edition, you’ll learn more about:
- DeepSeek’s new mathematical model
- Meta’s latest ChatGPT competitor
- A new approach to generating extended-length video
- AI shopping gets online payment giants’ blessing
But first:
Google Labs uses Gemini to create Little Language Lessons
Google engineers developed three experimental language learning tools powered by the Gemini API. The “Little Language Lessons” collection includes Tiny Lesson, which provides situation-specific vocabulary and phrases; Slang Hang, which generates authentic conversations between native speakers; and Word Cam, which uses image recognition to identify and translate objects in photos. Each experiment uses carefully crafted prompts to generate structured JSON outputs that deliver personalized, contextual language learning experiences. The tools demonstrate how AI can adapt to learners’ specific contexts, making language acquisition more natural and relevant than traditional methods. (Google)
Cognition AI’s DeepWiki offers free explanation of GitHub repositories
DeepWiki provides an instant way to understand unfamiliar codebases by automatically generating architecture diagrams, documentation, and source code links for public GitHub repositories. Users can access the tool by simply replacing “github.com” with “deepwiki.com” in any repository URL, with no installation required. The platform, powered by Devin Search, uses AI to create visual architecture maps, project summaries, technology stack breakdowns, and interactive file explorers that make complex codebases more approachable. DeepWiki’s conversational interface allows developers to ask specific questions about the code and receive context-grounded answers through its underlying DeepResearch agent. The service is free for public repositories, with support for private repositories available through authentication. (DeepWiki and Devin)
DeepSeek introduces new open model for mathematical theorem proving
DeepSeek-Prover-V2 is an open-weights large language model specifically designed for formal proofs in Lean 4. The model employs a novel recursive theorem-proving pipeline that uses DeepSeek-V3 to decompose complex mathematical problems into manageable subgoals while simultaneously formalizing these steps. After creating synthetic cold-start data by combining formal proofs with chain-of-thought reasoning, the team applied reinforcement learning to enhance the model’s ability to bridge informal reasoning with formal proof construction. The 671 billion parameter version achieves state-of-the-art performance with an 88.9 percent pass ratio on the MiniF2F-test benchmark and successfully solves 49 problems from PutnamBench. The researchers also introduced ProverBench, a new benchmark of 325 formalized problems from high school competitions and undergraduate-level mathematics. (GitHub)
Meta launches standalone AI assistant app powered by Llama 4
Meta AI, a competitor to ChatGPT and similar apps, remembers user preferences and maintains conversation context across interactions. The app enables voice conversations with natural dialogue capabilities, a discover feed for sharing AI-generated content, and integration with Meta’s existing AI features like image generation. Meta AI now serves as the companion app for Ray-Ban Meta glasses and connects with meta.ai on the web, allowing users to continue conversations across devices. The app is available now on iOS and Android, with voice features initially accessible in the US, Canada, Australia, and New Zealand. (Facebook)
SkyReels-V2 introduces infinite-length film generation
SkyworkAI unveiled SkyReels-V2, a new video generation model that enables extended-length film creation while maintaining visual quality and cinematic control. The model addresses key limitations in existing video generation systems by combining a multi-modal large language model with multi-stage pretraining, reinforcement learning, and a novel diffusion forcing framework. The researchers also developed SkyCaptioner-V1, a specialized video captioning system that accurately labels training data with detailed shot language and cinematic descriptions. Their approach uses motion-specific reinforcement learning to enhance dynamic movement quality and implements a diffusion forcing framework that enables generation of videos of unlimited length. Experiments show the model outperforms other open-source alternatives and enables applications including story generation, image-to-video synthesis, and camera direction. The team has made all code and models publicly available. (arXiv and GitHub)
Payment giants Visa, Mastercard, and PayPal race to enable AI shopping agents
Visa, Mastercard, and PayPal announced plans to deploy agentic commerce capabilities that will allow AI agents to complete purchases on behalf of consumers. The companies are integrating payment functionality into AI chatbots through partnerships with firms like Anthropic, Microsoft, and OpenAI, with rollouts expected in the coming quarters. Visa and Mastercard’s approaches rely on tokenization — creating secure digital payment credentials with spending limits that consumers can control — while PayPal offers developers API access tokens to integrate with its platform. Industry experts describe this shift as transformative, potentially shaping how consumers discover products and complete purchases while reducing return rates and improving shopping efficiency. (PYMNTS)
Still want to know more about what matters in AI right now?
Read last week’s issue of The Batch for in-depth analysis of news and research.
Last week, Andrew Ng highlighted an inspiring story of a high school basketball coach who learned to code and went on to teach computer science, emphasizing how AI helped scale K–12 education by empowering both students and teachers.
“Agentic workflows can automate a lot of teachers’ repetitive tasks. For example, when designing a curriculum, it’s time-consuming to align the content to educational standards (such as the Common Core in the United States, or the AP CS standard for many CS classes). Having an AI system carry out tasks like these is already proving helpful for teachers.”
Read Andrew’s full letter here.
Other top AI news and research stories we covered in depth: OpenAI launched API access to GPT Image 1, the image generator behind viral ChatGPT uploads; Google updated its AI-powered music generation tools, targeting professional musicians and creators; CB Insights’ Top 100 AI Startups list identified emerging players focused on AI agents and infrastructure; and researchers showed how large language models can improve shopping recommendations by inferring customer preferences from natural language input.