Morning Brief: Tuesday, September 22

One hundred seventeen feeds. Two weeks. 4,714 items reduced to what follows. (what we track, how we crawl, subscribe)

Tuesday is a model-launch pileup: Xiaomi's MiMo-V2.6-Pro claims the open-weights lead at a fraction of frontier training cost, xAI's Grok 4.7 arrives with multi-hour endurance claims it doesn't fully back up on launch day, and Meta's Muse agent draws both consumer adoption and platform pushback within the same news cycle.

A quieter connective thread: OpenAI's new math-advisory-group post shares its exact title with a same-day Terence Tao blog entry, suggesting he is part of it — which undercuts yesterday's read that Tao had moved on from AI-and-math commentary to routine coursework notes.

Top (5-7 min)

MiMo-V2.6-Pro 1T-A42B: the new top Open Weights model, trained for $3M: Latent Space, 2026-09-22. Xiaomi's MiMo-V2.6-Pro is framed as the new open-weights leader, a 1T-parameter (42B active) MoE reportedly trained for roughly $3M.
Grok 4.7: Hacker News, 2026-09-21. xAI's latest release, pitched around working autonomously for hours at a stretch.
Meta's Muse is outpacing ChatGPT's early mobile launch: TechCrunch, 2026-09-21. Meta's new personal AI agent is beating ChatGPT's download curve in its first days.
Meta's AI agent has been blocked from using Amazon.com: TechCrunch, 2026-09-21. Amazon shut Muse out the same week it's outpacing ChatGPT downloads.
OpenAI forms math advisory group as its AI resolves more than 100 open problems: TechCrunch, 2026-09-21. Follows a run of OpenAI-model claims on formal math problem-solving.
Value-Sensitive Delegation in Everyday AI Agent Use: Evidence from OpenClaw: arXiv, 2026-09-21. Field evidence on how people decide what to hand off to an AI agent versus keep for themselves.
Jev: System One models for Prod, not God — with Diogo Almeida, CEO, TypeSafe AI: Latent Space, 2026-09-21. Interview with the creator of the "Jev" small decision-model format that's been spawning HN forks all week.

Themes this week

MiMo-V2.6-Pro claims the open-weights lead: Latent Space: MiMo-V2.6-Pro 1T-A42B, trained for $3M (Tue), HN: MiMo v2.6 launch (Mon), Vercel: MiMo V2.6 models now available on AI Gateway (Mon).
Grok 4.7 ships with endurance claims that don't fully hold up: HN: Grok 4.7 (Mon), The New Stack: Grok 4.7 was built to work for hours. It still fails most of the time. (Mon), Vercel: Grok 4.7 now available and 40% off (Mon).
The "Jev" decision-model family keeps expanding: Latent Space: Jev, System One models for Prod not God (Mon), Simon Willison: Jev introduces a new shape of LLM (Mon), The New Stack: Your AI agent is burning tokens on choices that don't need words (Mon).
Meta's Muse draws adoption and platform pushback at once: TechCrunch: Muse is outpacing ChatGPT's early mobile launch (Mon), TechCrunch: Muse blocked from using Amazon.com (Mon), Slashdot: Muse beats ChatGPT in downloads, and gets blocked by Amazon (Tue).

Scan (15 min)

Tuesday feeds
- Can gzip be a language model?, Hacker News, 09-22
- Study: Young users (9 to 18Y) ditch Google for AI, with unknown consequences, Hacker News, 09-22
- Tiny Scratch-Built Cyberdeck is in Mint Condition, Hackaday, 09-22
- Cheap and easy throwaway VPN server, Lobsters, 09-22
- Jun Kim, oMLX creator and maintainer, joins Hugging Face, Hugging Face Blog, 09-22
- Fixing agent memory, InfoWorld, 09-22
- The agent coordination protocol hiding in plain sight: GitHub issues, InfoWorld, 09-22
Monday carryover
- Transformers Explained Visually, Hacker News, 09-21
- Igalia celebrates "Twenty-Five Years Upstream", LWN, 09-21
- Claude Code now also accepts instructions in OpenAI's Agents.md format, InfoWorld, 09-21
- AWS open-sources an AI agent it says is 45% cheaper than Claude Code and Codex, The New Stack, 09-21
- New npm malware finds a way around install script defenses, InfoWorld, 09-21

Tail

MiMo's $3M training-cost claim is single-sourced: Only Xiaomi's own announcement and downstream analysis (Latent Space, artificialanalysis.ai) are in the warehouse; no independent verification or skepticism has surfaced yet.
Muse's growth and its Amazon block are same-day, same-outlet stories: TechCrunch ran both "outpacing ChatGPT" and "blocked from Amazon.com" hours apart — an unresolved tension between adoption and platform friction that no other tracked outlet has picked up yet.
Grok 4.7's launch-day coverage argues with itself: The New Stack's stamina piece runs the same day as xAI's announcement and directly undercuts the "built to work for hours" framing — a same-day rebuttal rather than the usual multi-day lag.

Feed silences (>72h since last item)

Sources that publish frequently but have gone quiet:

Neel Nanda (399 days) — last item 2025-08-19.
Brendan Gregg (227 days) — last item 2026-02-07.
Spritely Institute (132 days) — last item 2026-05-13.
Andy Wingo (129 days) — last item 2026-05-16.
Aphyr/Jepsen (102 days) — last item 2026-06-12.
Typst (99 days) — last item 2026-06-15.
Eugene Yan (93 days) — last item 2026-06-21.
Lilian Weng (80 days) — last item 2026-07-04.
Andrej Bauer (73 days) — last item 2026-07-11.
Stephen Wolfram (63 days) — last item 2026-07-21.
Julia Evans (63 days) — last item 2026-07-21.
Marc Brooker (55 days) — last item 2026-07-29.
Antithesis (35 days) — last item 2026-08-18.
Bunnie Studios (23 days) — last item 2026-08-30.
METR (22 days) — last item 2026-08-31.
deepmind-blog (21 days) — last item 2026-09-01.
Vicki Boykis (21 days) — last item 2026-09-01.
GitHub Engineering (20 days) — last item 2026-09-02.
Fly.io (19 days) — last item 2026-09-03.

Build provenance