Meta released Muse Glimmer, a small fully open-weight model designed to run AI agents on-device, claiming it outperforms similarly sized Gemma4 and Qwen3.6 models across agentic, coding, and reasoning tests. Meta also said Muse Spark 1.2 will become open-weight soon.
xAI launched Grok Bot in beta as an iMessage-style chat where autonomous bots have dedicated computers, coordinate in individual or group chats, access websites and everyday apps, learn workflows, spawn specialist agents, and continue working 24/7. Access initially covers SuperGrok Heavy and top Cursor tiers across iPhone, Mac, Windows, and Linux, with broader availability promised.
Mark Zuckerberg’s essay argues that United States AI labs face disadvantages because of restrictions on training data, with high-quality code especially scarce. The discussion connects Meta’s Muse Code data strategy with the need for richer records of coding agents’ reasoning, mistakes, and repairs beyond polished public repositories.
Anthropic's unreleased Claude model, orchestrating 60 parallel subagents and 650 failed approaches, reportedly achieved the largest single improvement in 160 years on the Riemann Hypothesis. The resulting proof was verified in Lean 4, reviewed by two external mathematicians, and accompanied by public code and logs.
Researchers recovered proprietary reasoning from encrypted chain-of-thought traces returned by Anthropic, OpenAI, and Google. By replaying a frontier model’s trace into a weaker sibling and jailbreak-ing the weaker model, they extracted plaintext reasoning—including secrets—without directly attacking the stronger model or triggering anti-distillation safeguards.
Sign in to read the full issue and every issue in the archive. Free for 14 days.
Anything about the issue as a whole — a story that should have been here, a category that is off, two entries that are really one event, or how it reads.