Updates for Tuesday, March 17, 2026

Nvidia CEO Jensen Huang has projected an astounding $1 trillion worth of orders for the company's new Blackwell and Vera Rubin chips. This ambitious forecast underscores the immense anticipated demand for Nvidia's next-generation AI infrastructure, signaling a massive investment wave in artificial intelligence technologies. If realized, this would solidify Nvidia's dominant position in the AI hardware market and significantly impact the global tech economy.

OpenAI is reportedly expanding its government footprint by partnering with AWS to sell its AI systems to the U.S. government. This collaboration will cover both classified and unclassified work, significantly broadening OpenAI's engagement beyond its recent Pentagon deal. The move indicates a deeper integration of advanced AI into critical government operations and national security.

Senator Elizabeth Warren is pressing the Pentagon regarding its decision to grant xAI, Elon Musk's artificial intelligence company, access to classified networks. Warren expressed significant concerns, citing xAI's controversial chatbot, Grok, which has reportedly produced harmful outputs for users. She argues that this access poses a potential national security risk, prompting a high-level inquiry into the Pentagon's judgment.

Encyclopedia Britannica and Merriam-Webster have filed a lawsuit against OpenAI, alleging widespread copyright infringement. The dictionary publishers claim OpenAI illegally used nearly 100,000 of their copyrighted articles to train its large language models. This legal action highlights ongoing concerns about the unauthorized use of published content for AI development.

Mistral AI, a leading developer of open-source large language models, has announced a strategic partnership with NVIDIA. This collaboration aims to significantly accelerate the development and deployment of 'open frontier models,' leveraging NVIDIA's advanced AI hardware and software. The partnership is expected to boost the capabilities and accessibility of cutting-edge open-source AI.

DashClaw is a new tool that intercepts AI agent actions before they execute, addressing the problem of discovering agent activities only after they've occurred. It sits between agents and their actions, enforcing YAML-defined guard policies and allowing for human review and decision-making. This provides a crucial layer of control, enabling users to prevent unwanted actions, provide context, and record decisions for safer AI agent deployment.

MUP (Model UI Protocol) enables embedding interactive user interfaces directly within LLM chat, aiming to make powerful "agentic AI" accessible beyond text commands and developer tools. Each MUP is a single .html file, allowing functions to be triggered by both user clicks and LLM calls, with real-time interaction visibility for both parties. This protocol seeks to democratize agentic AI by providing a more intuitive and interactive user experience.

Oh-my-agent introduces a structural protocol to address common limitations of AI agents in real-world software development. It aims to mitigate issues like hallucinations, task drifting, and repetitive errors that arise from vague or bloated prompts and missing critical project details. By providing a more structured approach, this tool seeks to make AI agents more reliable and effective for practical coding tasks.

A comprehensive benchmark was conducted on 15 small language models (SLMs) across 9 diverse tasks, including classification, information extraction, and QA. The study aimed to provide data-driven insights to help developers choose the optimal base model for fine-tuning, addressing the challenge of selecting from numerous SLM options like Qwen3, Llama 3.2, and Gemma 3. This research offers practical guidance for fine-tuning efforts, moving beyond intuition to systematic evaluation.

Researchers compressed MLP layers of six LLMs without quantization, discovering that models degrade differently and some are significantly more compressible than others. For example, Gemma 2B retained 92% accuracy at 14% compression, while Llama 3.1 8B dropped to 85% at the same level. This surprising finding also revealed that initial perplexity improvements did not correlate with downstream benchmark performance after compression.

mlx-tune is a new Python library enabling native fine-tuning of Large Language Models (LLMs) and Vision-Language Models (VLMs) directly on Apple Silicon using Apple's MLX framework. It supports various advanced training methods like SFT, DPO, and GRPO, along with LoRA/QLoRA and GGUF export. The library's API mirrors popular frameworks like Unsloth/TRL, allowing developers to easily adapt existing training scripts for Mac, making LLM fine-tuning more accessible to Apple users.

mlx-tune is an open-source library that enables native fine-tuning of LLMs on Apple Silicon Macs, supporting various methods including SFT, DPO, and Vision. It features an Unsloth-compatible API, allowing developers to prototype training runs locally on their Macs to save on cloud GPU costs. This tool significantly streamlines the development workflow for Mac users in the LLM space.

Local LLMs, like Qwen 8B and 4B, have significantly improved their browser automation capabilities by shifting from upfront, full-task planning to a stepwise, replanning approach. This method allows the models to adapt to unexpected page states by planning one action at a time, making them much more robust and effective for real-world browser tasks compared to rigid, pre-conceived plans.

Covenant AI has introduced Covenant-72B, marking it as the largest model to date trained on decentralized, permissionless GPU nodes. This achievement was made possible by their innovative SparseLoco method, which builds on DiLoCo. SparseLoco significantly reduces communication overhead and synchronization frequency, utilizing a local AdamW optimizer and aggressive top-K sparsification to overcome bandwidth bottlenecks inherent in distributed training.

A man reportedly utilized ChatGPT to develop a personalized cancer vaccine for his dog, showcasing an unconventional and innovative application of AI. This case highlights the potential of large language models to assist in complex problem-solving and research, even in specialized fields like biotechnology and personalized medicine. While an anecdotal success, it suggests future possibilities for AI-driven solutions in healthcare.

Nvidia has launched NemoClaw, an open enterprise AI agent platform built upon the existing OpenClaw framework. This new platform is specifically designed to address and resolve Nvidia's primary security concerns within its AI offerings. By providing an open and secure solution, NemoClaw aims to accelerate the adoption of enterprise AI agents.

Memories.ai is developing a large visual memory model designed to index and retrieve video-recorded experiences. This technology aims to serve as a crucial "visual memory layer" for physical AI applications, particularly in wearables and robotics. It promises to enhance these systems' ability to learn from and recall past visual data.

Nvidia's new DLSS 5 leverages generative AI and structured graphics data to significantly enhance photorealism in video games. This advancement aims to make game visuals more lifelike than ever before. CEO Jensen Huang indicates that this generative AI approach has potential applications far beyond the gaming industry, suggesting a broader impact on various sectors.

HorizonMath is a new benchmark introduced to measure AI's progress toward mathematical discovery, specifically its ability to tackle important, unsolved problems. It features over 100 predominantly unsolved problems spanning eight domains in computational and applied mathematics. Coupled with an open-source automated verification framework, this benchmark aims to explore whether large language models can perform novel mathematical research.