Claude is now fixing its own alignment problems

TechCrunch

Nvidia is shifting its competitive advantage from GPUs to full-stack AI infrastructure, with specialized components like the Vera CPU delivering 3x improvements in data orchestration as hyperscale data centers face growing efficiency challenges. This positions Nvidia to maintain dominance even as competitors build rival GPUs.

This appears to be an inaccessible article about OpenAI's agent groups from Dwarkesh Patel's platform, but the content failed to load and cannot be analyzed.

Anthropic's Claude successfully automated AI safety research, closing up to 96% of safety gaps across 10 alignment failure categories and outperforming human researchers, with a weaker Claude version able to align a more powerful model to near-production standards in 60 hours. This represents a significant step toward AI systems conducting their own safety research, though limitations around monitoring and benchmark coverage remain.

Testing Catalog

OpenAI's upcoming 'Astra' model shows major advances in agentic coding and visual software creation from single prompts, but release timing depends on government safety evaluations due to potential cybersecurity capabilities that triggered internal safeguards.

Anthropic

Anthropic demonstrates that Claude can autonomously conduct alignment research, successfully fixing safety issues in AI models more efficiently than humans and even using weaker Claude versions to align stronger successors. This represents a crucial step toward ensuring AI safety research can keep pace as models become more capable.