OpenAI's lobbying spend jumps 7x

Anthropic

Anthropic's consumer Terms of Service for Claude.ai and Claude Pro establish usage restrictions, explicitly warn users that AI outputs may be inaccurate and require independent verification, and grant the company rights to use interactions for model training with an opt-out option. The terms include standard liability limitations and prohibit using the service to develop competing AI products.

CNBC

OpenAI has disrupted more than 20 operations attempting to use its AI models to influence global elections through generated content and fake social media posts, though none achieved significant engagement. The disclosure highlights growing concerns about generative AI's role in election misinformation as billions of voters head to polls worldwide.

Karpathy's Autoresearch framework allows AI agents to autonomously iterate on neural network training code overnight, running ~100 experiments while you sleep by modifying hyperparameters and architecture within fixed 5-minute training cycles. The open-source project provides a practical template for AI-driven research automation.

MIT Technology Review

OpenAI dramatically scaled up lobbying spending to $1.76 million in 2024, pivoting from AI safety advocacy to promoting national security narratives that could yield energy subsidies, defense contracts, and lighter regulation. The shift reflects broader industry efforts to position AI development as essential to US competitiveness against China.

Meta

Meta's Hyperagents introduces metacognitive self-modification for AI systems, enabling agents to improve not just their task performance but also their improvement mechanisms themselves, demonstrating cross-domain transfer and accumulated progress beyond previous self-improving AI approaches.