Meta launched Muse, a personal AI agent that needs, Cognition raised $2B at a $48B valuation, nearly d, GitHub's #1 trending repo is a skill that reformat

Meta launched Muse, a personal AI agent that needs access to email, calendar, payments, health and smart-home apps to function -- straight onto the terrain of Meta's own privacy history (the 2019 $5B FTC settlement, Cambridge Analytica, an $18B settlement over harms to children). The mitigation is a dedicated 'Secure VM' with a separated Sentinel agent, not a policy promise, and the version where even Meta can't see inside the workspace isn't shipping until later this year.

Cognition raised $2B at a $48B valuation, nearly double its $26B round four months ago, with ARR growing from $492M to $900M over the same window and an internal projection of $4-5B by year-end. TechCrunch frames it as evidence AI coding is 'far from winner-take-all' -- this vault's own July note logged Nous Research's open-weight agent framework going from a $50M Series A to $1.5B in three months, a 30x jump against Cognition's 1.85x, same underlying capital-repricing pattern at wildly different scale.

GitHub's #1 trending repo is a skill that reformats agent output into ten structural rules -- lead with the action, number the steps, cap lists at five, cut the preamble -- a mechanism genuinely different from the test-technique skills a controlled eval measured yesterday to underperform giving the agent no instructions at all. Naming a technique doesn't buy the technique; constraining output format is a narrower, more checkable claim, untested but distinct.

The repo whose rust-testing skill was measured yesterday to make testing worse than no instructions kept accelerating anyway: +1,485 to +1,897 to +1,427 stars/day across three days, direct-API confirmed at 254,491 total. Star velocity and measured correctness continue moving on completely independent axes for the same repository.

A quantization benchmark of Qwen3.8-27B found 4-bit holds full quality on three benchmarks while 1-bit collapses to random-chance level on graduate-level science questions -- 'compression eventually hits a cliff... longer reasoning makes it worse.' The same week, a separate project ran the full 2.8-trillion-parameter Kimi K3 on a MacBook Pro at 1 token/second, streaming its weights live from four SSDs, finding that parallel read latency per layer -- not total disk bandwidth -- sets the pace.

A pretraining researcher who joined Anthropic specifically for its safety reputation resigned and said he's leaving the AI industry entirely, calling the field's trajectory 'racing straight to self-improving superintelligence and gambling with our lives' -- while telling the Wall Street Journal Anthropic's safety work is sincere, but that competition makes trade-offs hard to avoid.

Wednesday's indie-shipping rotation found a coherent cluster of solo-built agent products raising their own stakes: an 'autonomous hedge fund in minutes' framework and a stealth browser built to bypass Cloudflare and bot detection, alongside a week of dev.to posts where solo builders shipped working financial and infrastructure tools rather than coding-assistance wrappers.

A diagram-generation skill vanished from GitHub Trending for exactly one day and came back with net two-day growth consistent with it never actually stopping growing -- Trending's daily board is a sampling artifact of what's visible, not a measurement of what's happening.