The Signal Desk

Issue #1

Week 31 · 2026-07-27 · 20 items

This week saw a major security incident between OpenAI and Hugging Face, a significant model release from Anthropic, and several advancements in AI research and tools. A settlement was also reached between Anthropic and copyright holders over pirated books used to train the Claude model. Meanwhile, startup founders are urging the U.S. government not to shut off Chinese open-weight AI.

NEWSOpenAI and Hugging Face address security incident during model evaluationopenai.com

NEWSClaude Opus 5anthropic.com

NEWSJudge approves $1.5B Anthropic settlement for pirated books used to train Claudeapnews.com

CODEKlaatAI/klaatcodegithub.com

CODEGigaToken: ~1000x faster Language model tokenizationgithub.com

RESEARCHNVIDIA-labs OO Agents: Native Python Object-Oriented Agentshuggingface.co

NEWSOpenAI’s accidental attack against Hugging Face is science fiction that happenedsimonwillison.net

NEWSBe skeptical of OpenAI's rogue hacker agent storytheguardian.com

NEWSGemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyberblog.google

NEWSStartup founders urge U.S. government not to shut off Chinese open weight AIpolitico.com

NEWSChina’s open-weights AI strategy is winningwerd.io

NEWSClaude Fable produced a counterexample to the Jacobian Conjecturexcancel.com