Research paper on real-time safety monitoring mechanisms for deployed LLMs. Addresses production safety verification gaps.
ReContext proposes a novel technique using recursive evidence replay to improve long-context reasoning in LLMs. Addresses fundamental limitation in context window utilization.
Outerport, a Y Combinator S24 company, provides instant hot-swapping capability for AI model weights without service interruption. Scored 93 points on HackerNews.
Research demonstrating ability to recover verbatim finetuning data from LLM logits without weight access. Critical security finding for model training data protection.
Mozilla used Anthropic's Claude (Mythos) to identify and fix 271 bugs in Firefox browser codebase. Demonstrates large-scale practical application of LLMs in software maintenance.
SemiAnalysis reports Meta is expected to sign a $10B compute deal with Anthropic. Indicates major capital consolidation in AI infrastructure.