The item most operators should act on today: OpenAI documents a separate Codex training setting that your ChatGPT opt-out does not change, and reports of it being silently re-enabled are unconfirmed. Also this week: Anthropic says its own pace of change may jump, Google plans default-off Gemini for school accounts, Apple sets an 18+ floor for dating-style apps, a Siri developer preview, and one research claim to ignore for now.
Three terms in this digest. Codex is OpenAI's AI coding assistant. A training setting decides whether OpenAI may use your conversations and code to improve its future models. The Responsible Scaling Policy is Anthropic's own published safety framework: it says which extra safeguards apply once its models reach defined capability levels.
The short version: check where your team's AI tools may learn from your data, and ask your AI suppliers how much notice you get when their models change.
If you run sales, logistics or retail, start with the OpenAI and Anthropic items: one is about how your information may be used, the other about how quickly the AI systems you depend on could change. This digest covers AI news for November 3, 2026, and we have not tested any product named here.
How we label items: Signal means an actionable development, Watch means a credible change that is not yet broadly usable or sufficiently validated, and Noise until reviewed means a claim that has not been independently reviewed. Each number belongs to whoever published it, and we read each item at its primary source on November 3.
Signal: Anthropic's filing says it "may imminently meet" an acceleration threshold
A TechCrunch headline on November 2 said Anthropic is close to self-improving AI. The primary document is narrower. (The Responsible Scaling Policy is Anthropic's own published safety framework, which sets extra safeguards once its models reach defined capability levels.) Anthropic's amended registration statement on Form S-1 (the document a company files with the US Securities and Exchange Commission before offering shares to the public), filed on November 2, 2026, says in a risk factor that internal estimates and expectations for future models, rather than current models, indicate it "may imminently meet" its Responsible Scaling Policy's definition of "dramatically accelerated progress attributable to AI". The next sentence says it has not designated any current model as exceeding that threshold. The filing also says Anthropic has in the past not identified a significant acceleration until after it occurred, and that its evidence of pace comes from lagging indicators. The phrase "RSI capability" is the press's, not the filing's, and the threshold is one defined in Anthropic's own policy, not an independent finding that self-improvement has begun. In the same filing Anthropic says it does not anticipate being able to safely use Claude to fully autonomously design and build successive generations of frontier models without meaningful human supervision or understanding.
Builder take: ask your model vendor how much warning you get before a model is changed or retired if it declares a threshold like this crossed, who re-tests your workflows afterwards, and whether that is in your contract. For operators: a faster-moving vendor is a reason to ask who pins model versions in the tools you buy.
Signal: OpenAI documents a separate Codex training setting; "silently re-enabled" is unconfirmed
A Hacker News thread says OpenAI is silently re-enabling model training on Codex chats. The thread is user reports. The original post points to the general "Improve the model for everyone" setting in ChatGPT, not to a Codex-specific one; some commenters report theirs changed, others say theirs stayed off, the poster says they cannot prove it, and we found no OpenAI statement. What OpenAI's Help Center says, read on November 3: on personal Free, Plus and Pro workspaces data sharing is on by default, and a separate article says turning off that setting or using the privacy portal is sufficient for new ChatGPT conversations and Codex tasks, with exceptions: feedback you give after opting out can send the whole conversation to training, and trusted-access programs such as Daybreak may use content even with the setting off. Codex also has a separate "Include environments" setting, which controls whether additional context from full Codex environments can be used for training, and changing the ChatGPT setting or using the privacy portal does not change it. The page states no default for that Codex setting. It also says content from ChatGPT Business, Enterprise, Edu and Healthcare workspaces, and API inputs and outputs, are not used for training by default.
Builder take: check both the ChatGPT setting and the separate Codex setting, and ask for the default of every training and retention toggle in writing, with a log of changes. For operators: if your developers use personal accounts for work, ask whether that is allowed, because the defaults differ from managed workspaces.
Watch: Google plans to put Gemini and Search behind an admin switch for K-12 students
A Google blog post dated November 3 by its education lead says Gemini, Gemini Notebook, Gemini in Classroom and Google Search will be "entirely off for K-12 student accounts until school administrators explicitly turn them on". Google says it is giving administrators controls and will share specifics with schools in the coming months before the changes roll out. This is an announced plan, not yet in effect. The post gives no effective date or age cutoffs, and the "early 2027" in some press reports is not in the text we read.
Builder take: if you sell into schools or any managed-account environment, ask the platform which AI features are tied to core features, who holds the switch, and whether turning AI off removes something else. For operators: default-off for a group is a policy choice, so ask your own vendors what their default is for staff accounts.
Signal: Apple sets an 18+ rating floor for apps that declare dating and relationship features
Apple Developer News of November 2 says that from November 16, 2026, App Store Connect's age-rating questionnaire adds a "Dating and Relationships" declaration, and all apps must complete it by January 12, 2027. The declaration covers features that enable dating or relationship connections between real people or with a virtual companion that engages a user in sexual or romantic content, and apps that declare it receive a minimum 18+ rating. It is a rating floor for apps that declare the feature, not a ban on AI companions for minors, and Apple notes local rules in Brazil.
Builder take: if your app offers a virtual companion designed or enabled to engage users in sexual or romantic content, review whether the new declaration applies, and check the intended features, supported interactions and age rating before January 12. For operators: no action unless your app includes dating or relationship features, including qualifying virtual companions.
Watch: Apple opens a developer preview of Siri AI, with limits
Apple Developer News of November 3 says Siri AI is open to all developers in a developer preview within the 27.3 betas of iOS, iPadOS, macOS and visionOS. The developer preview requires an Apple Intelligence-enabled device set to a supported language. Apps expose content and actions through App Intents, and new "skills" combine several actions into one experience. Apple says App Store submission opens in the spring, server-side model features are subject to daily usage limits with expanded access for a fee in the future, and the feature is unavailable in the EU on iOS, iPadOS and watchOS and for users under 13. Apple lists English, French, Japanese, Korean, Portuguese and Spanish, and says more detail on terms and usage policies will follow. A preview for developers is not the same as general availability for users.
Builder take: pick the three actions in your app you would let Siri trigger, and plan for what happens when a daily limit or the EU gap blocks them. For anything that changes an order or a payment, require the user to confirm and make sure a repeated request cannot place it twice; that is our recommendation, not an Apple requirement. For operators: if your staff or customers use an iPhone app of yours, ask your supplier whether they plan to expose actions to Siri.
Noise until reviewed: "LLMs can read and write neural network weights"
Noise here means noise as a buying signal, not as a research direction. MIT authors claim a model can edit a small model's weights and score far higher on maths, for example from 7.9% to 48.1% on MATH-500. A critique on Hacker News notes that only a few small experiments read raw weights with no tools, and that the large gain came from running Python code with up to 50 trial passes, a different technique from the headline. Simon Willison tested whether a local Qwen model could identify small game networks from their weights; it identified three of four and failed the fourth, and he does not read it as a reason to call source code obsolete. The authors say the results are "exciting but still limited", and we found no independent reproduction of the large gain. Interesting research, not a reason to change any purchase.
Do this today
The one action from this digest that takes ten minutes.
- Check both OpenAI training settings, the ChatGPT one and the separate Codex one, on every account your team uses for work.
- Ask each model vendor in writing how much warning you get before a model is changed or retired.
- If your product has a conversational character, check it against Apple's new dating and relationships declaration before January 12.
What to do this week
| News | Applies to | What to do now |
|---|---|---|
| Anthropic filing: "may imminently meet" a threshold | Teams building on a frontier model | Ask about release cadence, deprecation notice and version pinning in the contract. |
| Codex "Include environments" setting; silent re-enable unconfirmed | Anyone whose developers use Codex or ChatGPT accounts | Check both settings and ask for defaults and a change log in writing. |
| Google: Gemini and Search off for K-12 until admins enable | Education and managed-account products | Ask which AI features are tied to core features. |
| Apple: 18+ floor for dating and relationship declarations | Apps with dating or relationship features, including romantic virtual companions | Review your age rating before January 12, 2027. |
| Apple Siri AI developer preview | iOS app owners | Pick three actions to expose; plan for limits and the EU gap. |
| Weight-space programming claim | Nobody, yet | Treat as research; wait for independent reproduction on larger models. |
One habit for all six: ask who measured each number and what the filing, page or post actually says. Today's longer reads are our decode of the Postman agent-plugin report and our teardown of agent-desktop. For earlier context see the November 2 edition and our guide to production agent architecture.
Next edition: Signal / Noise, November 4.
Sources
- SEC EDGAR: Anthropic PBC Form S-1/A, filed November 2, 2026, read November 3, 2026
- OpenAI Help Center: Data controls in ChatGPT, read November 3, 2026
- OpenAI Help Center: How your data is used to improve model performance, read November 3, 2026
- Hacker News thread on Codex training settings, read November 3, 2026
- Google Blog, November 3, 2026: AI and learning updates for education
- Apple Developer News, November 2, 2026: new age rating content descriptors in App Store Connect
- Apple Developer News, November 3, 2026: Siri AI is now open to all developers
- Takehi, Teoh, Zhang and Khattab, November 2, 2026: Sparks of Weight-Space Programming
- Matthew Honnibal, November 3, 2026: critique of the weight-space post
- Simon Willison, November 3, 2026: Qwen can interpret model weights