AI Daily — May 29, 2026
2026-05-29
www.anthropic.com
Anthropic raises $65B Series H at ~$965B valuation
Anthropic has closed a $65 billion Series H funding round, pushing its post-money valuation to approximately $965 billion — making it one of the most highly valued private companies in history. The raise signals continued massive capital concentration in frontier AI development and positions Anthropic just below the trillion-dollar threshold ahead of a likely IPO.
www.anthropic.com
Introducing Claude Opus 4.8
Anthropic has released Claude Opus 4.8, a new iteration in its flagship Opus model line. Details from the announcement page were sparse in the feed, but given the version bump and Anthropic's concurrent Series H close, this release marks a significant model update worth tracking for capability and benchmark comparisons.
openai.com
OpenAI's Frontier Governance Framework
OpenAI has published its Frontier Governance Framework, detailing how its AI safety, security, and risk management practices align with emerging regulatory requirements from the EU AI Act and California AI legislation. The framework appears to be a structured attempt to map internal safety processes to external compliance obligations, which will be of interest to teams building on OpenAI APIs in regulated jurisdictions.
rss.arxiv.org
Mechanistic origins of catastrophic forgetting: why RL preserves circuits better than SFT
This paper moves beyond behavioral explanations for why RL fine-tuning causes less catastrophic forgetting than SFT, providing a mechanistic account at the level of attention head circuits. The authors introduce a 'differential circuit vulnerability' metric and show on Qwen2.5-3B-Instruct that RL updates preserve internal computational circuits more faithfully than SFT, correlating structural preservation with capability retention. This mechanistic framing has practical implications for fine-tuning strategy when preserving general capabilities is critical.