Vol. I · No. 199WED, NOV 4, 2026
Section · The Brief

Daily Brief

Google ties Search to AI for schools as Astra's weight-reading claims face a local-Qwen rebuttal.

Today in 30 seconds

  1. 01Google will cut Search access for schools that disable its AI tools.
  2. 02Willison runs Qwen 3.8 27B locally and reproduces GPT-6 Astra's weight-reading trick.
  3. 03Users report OpenAI silently re-enabled model training on Codex chats; unconfirmed.
  4. 04OpenAI pitches ChatGPT and Codex for CAD, PCB design, and simulation work.
  5. 05DeepMind's Antigravity gets Cheminformatics Skills to search compounds and propose macrocycles.
  6. 06New arXiv study tests how durably misaligned behaviors can be suppressed during training.

The Lead

Google will let schools disable its AI tools, but any school that restricts AI altogether loses access to Google Search, because AI features are now built into the search engine (TechCrunch, NYT). The policy turns an "opt out of AI" control into an "opt out of Google" decision for districts. It lands alongside Hacker News reports that OpenAI re-enabled "Improve the model for everyone" on Codex accounts and that Postman auto-installs AI plugins. Vendors are making AI the default and making refusal costly or hard to maintain.

Top stories

01

Google will cut Search access for schools that disable its AI tools

Google will let schools turn off its AI tools, but schools that restrict AI altogether will no longer be able to use Google Search. The company says AI features are now integrated into the search engine, so the two can't be separated. The NYT framed it as Google turning off search for schools that reject AI.

Why it matters

Bundling AI into a core utility removes the practical option of refusing it. Every district with an AI-restriction policy now has to choose between that policy and the default search engine.

Read the source on TechCrunch AI ↗
02
Open WeightsResearch

Willison reproduces GPT-6 Astra's weight-reading trick with Qwen 3.8 27B on a laptop

The paper "Sparks of Weight-Space Programming" (Takehi, Teoh, Zhang, Khattab) claims frontier models, notably GPT-6 Astra, can read and modify the raw weights of a machine learning model, given as a text dump of floating-point numbers. Simon Willison ran one of the paper's examples and got Qwen 3.8 27B on his laptop to do the same. A separate post by Honnibal calls the "Astra can read and write weights to improve Qwen" claim misleading.

Why it matters

If a 27B open model reproduces the demo locally, the capability isn't gated to frontier systems, and the headline claim for Astra is much weaker. It also hands practitioners a cheap way to test the paper's examples themselves.

Read the source on Hacker News (AI) ↗
03

Hacker News users report OpenAI silently re-enabling model training on Codex chats

A poster says their company had disabled data sharing in Codex, and that they and their coworkers found "Improve the model for everyone" switched back on without anyone opting in. The poster admits it may be isolated and notes that similar reports have appeared before. This is a user report, not a confirmed OpenAI change.

Why it matters

Companies that rely on that toggle to protect IP shouldn't assume it persists. Audit it at chatgpt.com/settings/data-controls, and expect pressure on OpenAI for an explanation.

Read the source on Hacker News (AI) ↗
05

DeepMind's Antigravity gets Cheminformatics Skills for macrocycle design

Google DeepMind shows Google Antigravity paired with its Cheminformatics Skills, used to accelerate drug-discovery workflows, search compounds, and propose novel rigidified macrocycles.

Why it matters

Google is packaging domain tooling as "Skills" for its agent platform. That is the same playbook OpenAI is running for physical-world engineering, applied to drug discovery.

Read the source on Google DeepMind ↗

Patterns

AI defaults beat opt-outs

Google's Search-for-schools policy, the reported Codex training-setting reset, and a GitHub issue alleging Postman auto-installs AI plugins and tries to share API data all show opt-outs eroding.

Local open models undercut frontier claims

Willison's laptop Qwen 3.8 27B run and Honnibal's critique both challenge the framing of GPT-6 Astra's weight-reading as a frontier-only capability.

Agents move into physical-world R&D

OpenAI targets CAD and PCB work, DeepMind targets macrocycle design, and the AgentAM paper exposes laser powder bed fusion simulation to an LLM agent via MCP.

Signal vs noise

Signal · underreported

Misalignment-suppression durability is an underreported safety question, and the companion result that fine-tuning on data a model believes false can induce broad misalignment makes it more pressing for routine fine-tuning.

Noise · overhyped

The "Astra reads and writes weights" framing, which a local 27B model reproduces and which Honnibal calls misleading.

What to watch

Watch whether OpenAI confirms or explains the Codex data-sharing reset, and whether Google offers schools any route to Search without AI features.

Sources

Filter sources
  1. Story 5
  2. Story 4
  3. Story 2
  4. Pattern 2
  5. Pattern 1
  6. Pattern 1
  7. Story 3
  8. Story 1
  9. Story 6
  10. You don't need long context
  11. Pattern 2
  12. Pattern 3