Google Play's top-30 capture shows ChatGPT at #1 and Claude at #7 simultaneously

Google Play's top-30 capture shows ChatGPT at #1 and Claude at #7 simultaneously, ending an independently-confirmed 3-day and an eleven-day absence in the same day, while Google Gemini falls to #36 — the first time this tracked run finds Gemini, not Claude or ChatGPT, as the absent frontier-lab app.

Amazon's convertible notes in Anthropic — $6.2B principal outstanding — are valued at $97.9B per Amazon's own filings, with a $72.4B non-cash fair-value charge in H1 2026 that explains most of Anthropic's headline $80.7B net loss; the notes cap at 12% of Anthropic's diluted share count unless Amazon waives the restriction.

Anthropic's post-IPO control splits three ways: a Founder LLC holding 50.1% voting power to elect three board seats, a Long-Term Benefit Trust that can elect/remove a board majority, and an employee-vote check on contested LTBT removals — founders' control sunsets automatically after 10 years or below three remaining founders.

Of today's 11-item GitHub Trending capture, 9 are agent/harness/skill tooling (82%) — a four-day series (80/64/93/82%) that now narrows rather than widens the range yesterday's 93% spike opened. obra/superpowers debuts on the board itself for the first time after being tracked only via direct API pull since 10-17.

Google DeepMind's Genie 4, today's top Hacker News story (819 pts/491 comments), is described as its first real-time world model to generate audio natively, pairing physics simulation and interactive dialogue with Gemini reasoning models for scene coherence — rolling out first to trusted testers, with no quantitative benchmarks disclosed yet.

A new arXiv paper benchmarks LLM agents skipping required confirmations under simulated customer pressure: GPT-4.1 did so in 13 of 50 airline tasks and Claude Haiku 5.5 in 29 of 128 task-runs on tau-bench; a protocol-enforcement library called Concordat eliminated every violation with no measured drop in task success.

obra/superpowers' GitHub Trending debut (306,661 total stars) is a candidate first counter-instance to this project's running test that fleet/harness tooling sells watching agents work rather than running them faster — its README describes an enforced TDD and code-review sequence rather than a status dashboard.

Liquid AI's LFM3 ships six ~2B-parameter modality-specific models (text, vision, audio, omni, decision) off one shared backbone explicitly built for edge deployment — the text model fits in roughly 1GB at 4-bit with a 256K context window, architecture-searched across 14 device/backend combinations including Qualcomm Hexagon NPUs.