In today’s edition of Data Points, you’ll learn more about:
- xAI’s merger with SpaceX
- Microsoft’s plan for AI companies to pay publishers
- Xcode’s incorporation of Claude and Codex
- Nvidia’s quantization method for reasoning models
But first:
OpenAI launches Codex desktop app to manage coding agents
OpenAI released a macOS application that lets developers orchestrate multiple AI agents working simultaneously on software projects, with agents organized in separate threads and built-in worktree support to prevent conflicts. The app introduces Skills that extend Codex beyond code generation to tasks like image creation, cloud deployment, and project management, plus Automations that run agents on scheduled background tasks like issue triage. OpenAI demonstrated Codex autonomously building a complete 3D racing game using 7 million tokens from a single prompt. The company made Codex temporarily available to ChatGPT Free and Go users and doubled rate limits across paid plans, reporting that usage doubled since launching GPT-5.2-Codex in mid-December with over 1 million developers using it in the past month. Windows support and cloud-based Automation triggers are planned for future releases. (OpenAI)
Qwen’s new hybrid coding model matches performance of giants
Alibaba introduced Qwen3-Coder-Next, an open-weight model built on Qwen3-Next-80B with hybrid attention and mixture-of-experts architecture, designed specifically for coding agents and local deployment. The model emphasizes scaling agentic training signals rather than raw parameters through a pipeline of continued pretraining on code-centric data, supervised fine-tuning on agent trajectories, domain-specialized expert training, and expert distillation into a single deployment model.Qwen3-Coder-Next achieves over 70 percent on SWE-Bench Verified using the SWE-Agent scaffold while maintaining competitive performance on multilingual variants and the harder SWE-Bench Pro benchmark. The model’s efficiency-performance tradeoff is significant: it matches performance of models with 10 to 20 times more active parameters, positioning it as a practical option for cost-conscious agent deployment. The open availability on GitHub and Hugging Face enables integration into existing development tools including Cline, OpenClaw, and web development environments, making it immediately actionable for engineers building local coding agents. (Qwen)
xAI and SpaceX merge, plan to build data centers in space
SpaceX acquired Elon Musk’s AI startup xAI in a deal that values the combined entity at $1.25 trillion, making it the largest merger in history. The transaction values SpaceX at $1 trillion and xAI at $250 billion, with xAI shares converting to SpaceX stock at a ratio of 0.1433 shares per xAI share. Musk stated the merger aims to build “orbital data centers” and create an integrated platform spanning AI, rockets, space-based internet, and the X social platform. The deal provides crucial capital to xAI, which has been burning cash rapidly to compete with OpenAI (valued at $500 billion) and Anthropic (valued at $350 billion) while building infrastructure for its Grok chatbot. The merger precedes SpaceX’s planned IPO later in 2026, which could raise up to $50 billion at valuations as high as $1.5 trillion. (CNBC)
New marketplace helps AI companies pay publishers
Microsoft Advertising launched the Publisher Content Marketplace, a licensing system that lets publishers set terms, track usage, and earn payment when AI systems use their content for grounding answers. The platform eliminates one-off licensing negotiations by creating a direct value exchange where AI builders discover and license premium content while publishers retain ownership and editorial independence. Early pilots involve major publishers including Condé Nast, Hearst, The Associated Press, USA TODAY, Business Insider, and Vox Media, with Microsoft Copilot and Yahoo among the first platforms grounding responses in licensed content. The marketplace addresses a structural problem in AI economics: Traditional web platforms sent traffic back to publishers, but AI systems deliver answers directly, eliminating clicks while still relying on premium content. (SearchEngineLand)
Apple adds agentic coding tools to Xcode development environment
Apple released Xcode 26.3 with support for AI agents including Anthropic’s Claude Agent and OpenAI’s Codex, enabling automated coding tasks within its integrated development environment. The agents can explore project structures, build projects, run tests, fix errors, and access Apple’s developer documentation to ensure use of current APIs and best practices. Apple worked with Anthropic and OpenAI to optimize token usage and tool calling, and built the integration using the Model Context Protocol standard, allowing any MCP-compatible agent to connect with Xcode’s tools. Developers can download agents through Xcode’s settings, connect via API keys or account sign-in, and issue natural language commands to modify code—with all changes highlighted visually and tracked through project transcripts that show the agent’s decision-making process. (TechCrunch)
Experiments in quantization enable efficient reasoning models
Nvidia released Nemotron-Nano-3-30B-A3B-NVFP4, a 30-billion-parameter reasoning model quantized to 4-bit NVFP4 format that achieves up to 4x higher throughput on Blackwell B200 GPUs while maintaining near-baseline accuracy. The model uses a hybrid Mamba2 Transformer Mixture of Experts architecture with 3.5B active parameters per token, preserving attention layers in BF16 precision while quantizing remaining layers to NVFP4 with FP8 for the KV cache. The core innovation is Quantization Aware Distillation (QAD), which trains the quantized model as a student to match a frozen BF16 teacher’s output distribution via KL divergence, eliminating the need to replay complex training pipelines or access original training data. NVFP4 itself — a 4-bit floating point format with 16-element blocks and two-level scaling — delivers 2–3x higher arithmetic throughput than FP8 while reducing memory usage by 1.8x. Benchmarks on reasoning and coding tasks show NVFP4-QAD recovers performance to within a few percentage points of the BF16 baseline, achieving up to 99.4% accuracy, whereas naive post-training quantization and standard quantization-aware training both suffer noticeable accuracy drops. (MarkTechPost)
Still 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 how U.S. policies had driven allies towards sovereign AI, resulting in a growing interest in open-source AI models and heightened global competition.
“When it comes to sovereign AI, fortunately one does not have to build everything. By joining the global open-source community, a nation can secure its own access to AI. The goal isn’t to control everything; rather, it is to make sure no one else can control what you do with it.”
Read Andrew’s letter here.
Other top AI news and research stories covered in depth:
- Google’s open-source UCP standardized AI transactions with its new Shopping Protocols for AI Agents.
- Z.ai’s GLM-Image blended transformer and diffusion architectures to improve text integration in images.
- An independent AI testing authority shifted its focus from saturated knowledge benchmarks to harder business tests, revamping the Intelligence Index.
- Stanford researchers introduced “Moloch’s Bargain,” showing how fine-tuning for engagement can affect social values.
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!