Tabular Foundation Models (TFMs) are advanced AI designed specifically for handling tabular data, unlike conventional large language models (LLMs) that struggle with such data. TFMs treat numeric values accurately and seek to discern essential patterns within rows and columns, enhancing insight discovery in spreadsheets and structured datasets as they evolve.
At Kilo Code, engineers now spend only 1% of their time coding, relying heavily on AI agents. This shift raises questions about system safety and cost management. Companies like Replit and Symbotic adapt by implementing human oversight and managing AI expenditures carefully as multi-model support evolves to enhance efficiency and reduce wasted budgets.
GitHub has introduced stacked pull requests, a feature that allows developers to divide a large feature into smaller, ordered pull requests for easier review. This technique enhances review quality and speed, addressing issues arising from larger pull requests. As AI tools generate more code, efficient review processes become essential for maintaining productivity and quality.
In October 2025, Hurricane Melissa struck Jamaica, significantly impacting communities. The AI model WeatherNext, developed by Google’s DeepMind, accurately predicted the storm’s trajectory with 80% confidence, offering forecasters an extra day of warning. This breakthrough improves preparedness for extreme weather events, showcasing advancements in hurricane prediction accuracy.
OpenAI’s Astra model has solved ten longstanding mathematical and theoretical computer science problems, publishing machine-checkable proofs and certificates on GitHub. Key results include a non-sofic group construction and disproving Connes’s rigidity conjecture. Though significant, none of the findings have undergone peer review, raising concerns within the mathematics community regarding AI’s impact on research integrity.
BMW plans to deploy the Figure 03 humanoid after a successful pilot. Despite advancements in robotics and AI, manufacturing challenges persist, primarily due to high costs and financial viability. Specialized robots currently outperform humanoids in efficiency, leading to a critical need for reduced pricing and expanded production capabilities to achieve market competitiveness.
Visa leveraged Anthropic’s Claude Mythos to identify vulnerabilities in its extensive payment infrastructure, spanning over 200 countries and handling billions of transactions. The approach emphasized proactive security measures like the Visa Vulnerability Agentic Harness, prioritizing rapid adaptations to threats and recognizing the importance of supply chain security amid evolving AI risks.
General Motors’ autonomous driving division utilizes AI agents to optimize workflows, with engineers spending only 15% of their time coding. By automating tasks such as data analysis and problem triaging, GM has tripled the number of merged pull requests, leading to faster releases and fewer defects. This approach redefined workflows beyond simple coding assistance.
Moonshot AI’s Kimi K3 model is reportedly faster than competitors, raising concerns among US companies and officials about potential technology theft. While proponents argue that open-weight models promote innovation and affordability, critics caution against misuse. A divide exists between supporting open models and concerns over security and competition, influencing future AI development policies.
OpenAI has open-sourced its Codex Security command-line interface and software development kit under the Apache 2.0 license, allowing integration into development workflows. While teams can adapt the client code, the underlying AI scanner remains under OpenAI’s control, raising questions about reliability in CI processes for validating code changes.
Huawei’s release of openPangu-2.0-Pro, a 505-billion-parameter language model trained entirely on its Ascend 910B NPUs without Nvidia hardware, marks a significant milestone in AI. However, the training utilized components partially sourced from TSMC and Samsung, raising questions about the sustainability of a fully domestic hardware stack. The model’s architecture includes advanced techniques for efficient training and inference.
OpenAI has announced significant price reductions for its GPT-5.6 model family, with Luna’s rates dropping by 80%, now costing $0.20 per million input tokens. This change results from self-optimizations made by the Sol model, improving inference infrastructure efficiency. The adjustments enhance economic viability for high-volume applications, reflecting competitive market pressures.
South Korean industrial robots are generating real-world training data for NC AI’s World Foundation Model, aimed at overcoming the Sim-to-Real gap in autonomous robotics. NC AI partnered with Cmassrobotics to create a data loop that improves AI models using real factory data, enhancing efficiency and accuracy for manufacturing and logistics automation.
Researchers introduced DataFlow-Harness, an open-source framework that guides AI agents in creating structured data-processing workflows, addressing the “NL2Pipeline gap” between natural language tasks and production readiness. This approach ensures persistent, editable artifacts while enhancing accuracy and reducing API costs significantly, achieving a 93.3% success rate in managing complex data pipelines.
