Hosts: Lenar Kess, Damra Vol. Anthropic says its new Claude models carry a watermark inside the generated text itself, not in metadata — which means the signal survives copy-paste and travels into every corpus downstream. We work through what that actually buys, and what it costs. Then: OpenAI trains exploit development on purpose and gates it behind tiered access, NVIDIA helps arrange the money that buys NVIDIA chips, Muse Glimmer arrives under Apache 2.0, Bernie Sanders writes a letter while a think tank writes a mechanism, Linus Torvalds talks about review throughput, and a water-use paper meets a letter to Texas.Anthropic's announcement of in-text watermarking for the new Claude modelsPaul Graham's reaction to the watermark newsOpenAI on GPT-5.6-Cyber and the Daybreak Blue access tierEric Wallace on training exploit-development capability deliberatelyEthan Mollick on what tiered access impliesNVIDIA's financing platform announcementSimon Willison on Muse Glimmer's Apache 2.0 licenceArtificial Analysis's Intelligence Index placement for Muse GlimmerAndrew Curran on the Sanders letter and the ARI governance proposalLinus Torvalds on review throughput on the kernel mailing listCactus's Needle2 and the Ante vision connector releasesWater Research on data-center water footprint and siting
