Morning Brief: Monday, September 21

One hundred seventeen feeds. Two weeks. 4,902 items reduced to what follows. (what we track, how we crawl, subscribe)

Monday's headline story graduates the weekend's slowdown chatter from commentary to litigation: a lawsuit accuses Anthropic, OpenAI, SpaceX AI, and Google of an illegal agreement to slow AI development, landing a day after TechCrunch asked whether the industry was "really ready to slow down" and Trump's "AI Force" proposal drew industry head-scratching. A second, lighter thread: Hacker News keeps growing the open "Jev" small-model family, with two more entries riffing on last week's experiments.

The suit's defendant list is notable for pairing SpaceX AI with the three usual foundation-model labs, and it lands without a matching mainstream follow-up yet in the warehouse — worth checking whether coverage catches up tomorrow.

Top (5-7 min)

Lawsuit Says Anthropic, OpenAI, SpaceXAI And Google Made Illegal Agreement On AI Slowdown: Slashdot, 2026-09-21. Names all three major US foundation-model labs plus SpaceX AI as co-defendants, following a weekend of commentary on whether the industry is actually ready to slow down.
Heretic removes restrictions from language models: Hacker News, 2026-09-21. A tool for stripping refusal/safety training from open-weight models, framed as restoring base-model behavior rather than jailbreaking.
Kev: Tiny Jev-like family of decision models built on top of Qwen3.5: Hacker News, 2026-09-21. Latest HN entry riffing on the small "Jev" decision-model format, this one built on Qwen3.5.
Show HN: Mini-AGI – Dynamic continual learning model trained on 8GB VRAM: Hacker News, 2026-09-21. A from-scratch continual-learning model sized for a single consumer GPU, set against the trend of ever-larger frontier training runs.
OpenAI's cyber defense letter gets the diagnosis right and the prescription wrong: InfoWorld, 2026-09-21. Argues OpenAI correctly names AI-accelerated attack tooling as a real threat but proposes centralization that mainly benefits incumbents.
LEGIT: Credentialing Protocol for Trustworthy AI Agent Marketplaces: arXiv, 2026-09-21. Verifiable credentials for agents transacting in marketplaces, part of a same-day cluster of agent-governance papers.

Themes this week

AI-slowdown scrutiny escalates from op-ed to lawsuit: Slashdot: Lawsuit says Anthropic, OpenAI, SpaceX AI, Google made illegal slowdown agreement (Mon), TechCrunch: Is the AI industry really ready to slow down? (Sun), Slashdot: Tech industry scratches its head over Trump's AI Force proposal (Sun).
The open "Jev" model family keeps growing: HN: Jev-Leftpad (Mon), HN: Kev, tiny Jev-like models on Qwen3.5 (Mon), HN: I turned Jev into a (lousy) chatbot (Sun).
AI agent governance and credentialing papers cluster: arXiv: LEGIT, credentialing protocol for AI agent marketplaces (Mon), arXiv: Authorization Revocation for Long-Running AI Agents (Mon), arXiv: CIPL, channel-aware framework for recoverable privacy leakage in LLM agents (Mon).

Scan (15 min)

Monday feeds
- A KV cache you can fork, Planet Clojure, 09-21
- 20 approaches to writing better AI prompts, InfoWorld, 09-21
- What Sun got wrong, Lobsters, 09-21
- Why Backprop Goes Backward (2018), Hacker News, 09-21
- AI chatbots give wrong answers to financial queries 'most of the time', Hacker News, 09-21
- Grim Fandango Puzzle Document (1996) [pdf], Hacker News, 09-21
Sunday carryover
- Your AI agent failed. The model might not be the problem., The New Stack, 09-20
- World model companies are keeping a lot of secrets, TechCrunch, 09-20
- Deterministic Core, Non-Deterministic Shell, Lobsters, 09-20
- datasette-explain 0.2.2, Simon Willison, 09-20
- Kernel prepatch 7.3-rc4, LWN, 09-20

Tail

Tao switches from human-role debate to technical notes: After three straight posts this week on AI's role in mathematics, Monday's Tao entry is course notes on rearrangement-invariant spaces — a topic shift rather than a continuation, and possibly a sign the debate thread has run its course for now.
Lawsuit lacks mainstream corroboration in the warehouse: The Slashdot item is currently the only warehouse entry on the Anthropic/OpenAI/SpaceX AI/Google slowdown suit; no TechCrunch, Reuters, or Bloomberg pickup has appeared yet as of this build.
The "Jev" family has no canonical origin post in the tracked window: Entries have surfaced as individual HN show-and-tell posts (chatbot port, Qwen3.5 port, leftpad joke) across three days with no announcement or reference implementation post tying them together.

Feed silences (>72h since last item)

Sources that publish frequently but have gone quiet:

Neel Nanda (398 days) — last item 2025-08-19.
Brendan Gregg (226 days) — last item 2026-02-07.
Spritely Institute (131 days) — last item 2026-05-13.
Andy Wingo (128 days) — last item 2026-05-16.
Aphyr/Jepsen (101 days) — last item 2026-06-12.
Typst (98 days) — last item 2026-06-15.
Eugene Yan (92 days) — last item 2026-06-21.
Lilian Weng (79 days) — last item 2026-07-04.
Andrej Bauer (72 days) — last item 2026-07-11.
Stephen Wolfram (62 days) — last item 2026-07-21.
Julia Evans (62 days) — last item 2026-07-21.
Marc Brooker (54 days) — last item 2026-07-29.
Antithesis (34 days) — last item 2026-08-18.
Bunnie Studios (22 days) — last item 2026-08-30.
Microsoft Research (21 days) — last item 2026-08-31.
METR (21 days) — last item 2026-08-31.
deepmind-blog (20 days) — last item 2026-09-01.
Vicki Boykis (20 days) — last item 2026-09-01.
GitHub Engineering (19 days) — last item 2026-09-02.
Fly.io (18 days) — last item 2026-09-03.

Build provenance