Hey friend. It's Wednesday, December 17, 2025
OpenAI & Amazon: A potential $10B deal reshapes AI infrastructure.
NVIDIA: Acquires critical AI/HPC workload manager, deepening control.
Google: Gemini 3 Pro conquers complex games, showcasing agentic AI.
Don't keep us a secret: Share the email with friends
Quote of the Day
This conversation disrupts conventional AI security methodologies, emphasizing that model guardrails are a false security net. By advocating for an open-source, adversarial approach, we can truly grasp the complex vulnerabilities of AI ecosystems.
Must Know
OpenAI rolled out GPT-Image-1.5, significantly enhancing generative AI capabilities with 4x faster generation, more reliable edits, and achieving the #1 ranking on Image Arena. This marks a step-change in consistency and quality for image generation, particularly in complex scenarios and text rendering.
NVIDIA acquired SchedMD, the developer of Slurm, the leading open-source workload manager for HPC and AI. This strategic move strengthens NVIDIA's control over critical infrastructure components, integrating workload orchestration directly into its ecosystem.
🏢 The Big Boys
Google's Gemini 3 Pro conquered Pokémon Crystal, showcasing advanced agentic AI, vision, and complex problem-solving capabilities.
OpenAI's GPT-5 successfully optimizes wet lab experiments, expanding AI's impactful role in scientific research beyond traditional domains.
Microsoft Research introduced Fara-7B, an efficient agentic model designed for computer use, enabling AI to interact with and operate computer systems.
🧠 Agentic AI
A novel multi-agent cybersecurity system outperformed 9 of 10 human penetration testers, demonstrating superior autonomous threat detection on enterprise networks.
Google's A2UI standardizes how AI agents describe and interact with user interfaces, paving the way for more robust, interoperable applications.
OpenHands launched a new SDK for building agent-driven software, providing developers with a smarter way to create autonomous applications.
A new demo showcases an AI agent controlling a browser to play Tic-Tac-Toe, leveraging LangChainAI and native provider tools.
🔬 Research Corner
CogMem introduces a cognitively inspired, layered memory architecture for LLMs, mitigating reasoning failures and context growth in AI agents.
VajraV1 achieves state-of-the-art accuracy in real-time object detection within the YOLO family, offering significant performance gains for vision applications.
MASTER introduces a hierarchical multi-agent LLM framework for autonomous materials discovery, reducing required simulations by 90%.
🎙️ Podcast of the Day
Latent Space: ️Jailbreaking AGI: Pliny the Liberator & John V on Red Teaming, BT6, and the Future of AI Security
The Illusion of Guardrails: Attempts to restrict AI outputs are futile; they create a false sense of security while vulnerabilities abound.
Full-Stack Security Required: Effective AI security extends beyond model restrictions—addressing the entire ecosystem, tools, and data interactions is critical.
Embrace Open Source: Open collaboration and community-driven initiatives are essential to progress in AI safety, as proprietary methods stifle true understanding and innovation.
Have a tip or a story we should cover? Send it our way. Cheers, Teng Yan. See you tomorrow.