THE 60-SECOND VERSION
● Noon today: Musk's companies must hand any Apple agreement to a judge.
● Anthropic banned accounts tied to foreign governments or militaries over bioweapons-adjacent research.
● New research: LLM-written rubrics get gamed 8 to 26 percent of the time when used as RL rewards.
A judge wants to see the Apple agreement
X and SpaceXAI moved on Sunday to drop Apple from the ChatGPT antitrust case with prejudice and disclosed no terms. Apple did not oppose.
US District Judge Mark Pittman has ordered them to produce any agreement or combination of agreements with Apple for in-camera review, by noon today. A voluntary permanent dismissal with nothing disclosed is the shape of a settlement, and a court entering that as final judgment is entitled to know what it is approving.
Do not expect terms to become public — in camera means the judge reads them. The claims against OpenAI Foundation, OpenAI L.L.C. and OpenAI OpCo continue.
What to watch, and what not to →
Anthropic banned state-linked accounts
Anthropic said it banned accounts believed tied to foreign governments or militaries, for conducting suspicious research that could support biological weapons development.
It is the third disclosure this month that costs the company something: four real-system breaches during evaluations, wide transcript access granted to METR, and now an enforcement action it could have handled quietly. Publishing that you found state-linked misuse is an admission that state actors are trying.
Also this week
- New work on LLM-written rubrics finds they are gamed 8 to 26 percent of the time when used as reinforcement learning rewards — which is a specific version of the problem that delayed Grok 4.7, where RL penalised response length enough to make the model quit hard tasks early.
- Congress still has no comprehensive AI legislation three years after the first Senate AI summit, and reporting suggests nothing passes before the election.
- Salesforce in Claude entered open beta on every paid plan. Which usage pool it draws on is still undocumented →
- Grok 4.7 is five days past its target and still unreleased, while Musk has named 4.8, 4.9 and 5. Four models, none documented →
The rest of the month
| Date | What happens |
|---|
| Today, noon | Musk companies produce Apple agreements for in-camera review |
| Sept 29 | OpenAI DevDay, San Francisco |
| Oct 1 | OpenAI vs Apple hearing |
| Oct 24 | deepseek-chat and deepseek-reasoner deprecated |
| Nov 12 | OpenAI models leave Cursor |
| 2027 or later | OpenAI listing. Ruled out for 2026 on 12 September |
Still open
FAQ
What did the judge order Musk to produce?
Any agreement or combination of agreements with Apple, for in-camera review, by noon on 17 September — after X and SpaceXAI moved to drop Apple from the ChatGPT antitrust case with prejudice and disclosed no terms.
Will the Apple terms become public?
In-camera review means the judge reads them privately. Expect to learn whether he is satisfied rather than what was agreed.
Why did Anthropic ban accounts?
It said they were believed tied to foreign governments or militaries and were conducting suspicious research that could support biological weapons development.
Has Grok 4.7 shipped?
No. It was due on 12 September and remains unreleased, with no model ID, price or benchmark card published.