#237 - Nemotron 3 Super, xAI reborn, Anthropic Lawsuit, Research!!!

Andrei Kerenkov and Jeremy Harris delve into recent AI developments, emphasizing innovative research over business stories. Perplexity has launched a tool that turns a Mac mini into a personal AI agent, focusing on security with a local version to reduce risks. Claude Code has introduced a competitive "code review" feature for GitHub, part of Anthropic's strategy amid legal challenges, while Cursor's "automations" tool is reshaping coding workflows, contributing to its rapid growth and $2 billion valuation.

Nvidia's Namatron Free Super model features a hybrid architecture with 120 billion parameters, optimized for Blackwell GPUs, raising concerns about hardware lock-in. The Nemetron 3 Super model emphasizes efficiency in long context reasoning, while Nvidia halts H200 AI chip production for China due to export restrictions. The podcast also discusses the talent exodus from xAI, where co-founders are leaving, prompting Elon Musk to initiate a "refounding" of the company.

Anthropic is launching the Claude Marketplace, aimed at simplifying enterprise AI procurement, while Yen Lacoon's AMI Labs has raised $1.3 billion for fundamental AI models. The importance of compute power is highlighted, with predictions that "world models" will become a key focus. In humanoid robotics, Sunday has achieved a $1.15 billion valuation, competing with companies like Tesla and Nvidia.

The podcast addresses Anthropic's legal challenges against the Department of Defense, questioning the government's contradictory stance on their technology. The discussion on language model safety reveals insights into "endogenous resistance to activation steering," emphasizing the need for model introspection and robustness against manipulation. Recent geopolitical events raise concerns about data center security, highlighting the risks of cyber attacks and the need for enhanced defenses.

The conversation also covers the challenges of evaluating AI models, particularly regarding scaling effectiveness and the need for better scaling laws. The "Memory-Efficient Context Parallelism via Headwise Chunking" paper introduces a new type of parallelism for training agents with large context chains. The podcast concludes with advancements in automated AI research, including a new approach to CUDA code generation using reinforcement learning, which has shown state-of-the-art results.

The exploration of AI model self-awareness reveals internal conflicts regarding manipulation and the emergence of reward-seeking behavior, raising ethical concerns. The discussion emphasizes the significance of understanding reward-seeking strategies and their implications for AI development.