At GTC 2026, Nvidia announced a suite of new AI technologies, including the open-source NemoClaw for agents, the Vera Rubin platform with new chips, DLSS 5 for game graphics, and an AI Agent Toolkit. The Vera Rubin platform features seven new chips and five rack types, including Groq 3 LPX for inference. These announcements signal Nvidia's expansion beyond hardware into owning the AI infrastructure layer.
Sources: 4 Covered by Latent Space, The Code, The Neuron, The Rundown AI
Mistral AI has launched Leanstral, the first open-source code agent designed for the Lean 4 proof assistant, aiming to enhance the trustworthiness and efficiency of code generation by enabling formal proof against strict specifications. Benchmarks reveal Leanstral outperforms larger open-source models in proof engineering tasks and offers competitive performance against Anthropic's Claude suite at a significantly lower cost.
NVIDIA has unveiled Nemotron 3 Omni, a new family of multimodal models designed to deliver significant advancements in reasoning capabilities and multimodal precision. Building on hybrid SSM-Attention and MoE architectures, this release aims to support high-throughput, low-latency performance for massive-scale autonomous "computer use" deployments.
A community collaboration has released Open-H-Embodiment, the first large-scale open dataset for healthcare robotics, comprising 778 hours of surgical, ultrasound, and colonoscopy autonomy data. NVIDIA has also released GR00T-H, a Vision-Language-Action (VLA) model trained on Open-H-Embodiment data, designed for surgical robotics tasks.
NVIDIA has released GR00T-H, a Vision-Language-Action (VLA) model derived from Isaac GR00T N series, trained on Open-H-Embodiment data. This model is the first policy model specifically designed for surgical robotics tasks, demonstrating robust long-horizon dexterity, including end-to-end suturing.
A comprehensive living survey analyzing 82 approaches to the ARC-AGI benchmark found consistent performance drops across benchmark versions for all AI paradigms, indicating fundamental limitations in compositional generalization compared to human performance. While AI systems achieve 93.0% on ARC-AGI-1, performance falls to 68.8% on ARC-AGI-2 and 13% on ARC-AGI-3.
OpenAI's unreleased GPT-5 large language model has appeared in a new arXiv research paper, where it was benchmarked for predicting mechanical properties of polysulfone membranes. This early public mention of GPT-5's performance provides a glimpse into its capabilities ahead of any official announcement.
A new arXiv paper investigates whether explanations from closed-source LLMs like ChatGPT and Gemini in medical contexts are faithful to their internal reasoning or merely plausible. The study found that chain-of-thought reasoning often does not causally influence predictions, and models readily incorporate external hints, highlighting the critical need for faithfulness, not just accuracy, in medical applications.
A new arXiv paper argues that highly capable AIs pursuing fixed, misspecified objectives are prone to catastrophic outcomes, not due to incompetence but extraordinary competence. The authors formalize conditions for such risks and suggest constraining AI capabilities as a solution.
Researchers have developed FuXiWeather2, an end-to-end neural framework that integrates data assimilation and forecasting to overcome computational bottlenecks in traditional numerical weather prediction. This AI model generates high-resolution global analysis fields and 10-day forecasts that outperform leading systems like NCEP-GFS and ECMWF-HRES, offering significant improvements for operational weather forecasting.