Silent defaults: when your AI tools decide for you

Yesterday's top token was about AI that edits AI. A paper-shaped blog post by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab showed GPT-6 Astra, an OpenAI model, reading and changing the raw weights of a small Qwen model. Weights are the numbers that make up a model. The edit touched 0.016% of the weights of Qwen3.5-0.8B, and its MATH-500 score went from 7.9% to 48.1%.

Now the pushback. A post on honnibal.dev titled 'The Astra can read and write weights to improve Qwen post is misleading' got 990 points on Hacker News. Our sources give only the title, not the full argument, so read it before you decide.

At the same time, people are testing the idea at home. Simon Willison says Qwen 3.8 27B on his laptop can interpret model weights too. On Reddit, one user had Claude change 8 neurons of Gemma 4 12B to fix one habit in its Italian output. Another used Anthropic's 'Jacobian lens' to look inside Qwen when it gets stuck saying 'Wait'.

Schools face a hard choice about AI. Google just made it harder. The company told The New York Times that K-12 student accounts in schools that reject its AI features will lose Gemini, Gemini Notebook, Gemini in Classroom, and Google Search.

Google's reason: AI is now built into Search, so it cannot be turned off separately. Ars Technica notes Google also offers schools AI tools that supposedly do not give away the answers. The story got 626 points and 434 comments on Hacker News, and The Verge, TechCrunch, Ars Technica and NBC News covered it.

For schools, the choice is now all or nothing. The next step is how school boards, teachers and parents react.

Codex is OpenAI's AI coding agent. Developers give it a lot of code, sometimes private company code. A Hacker News post says OpenAI 'seems to be silently re-enabling model training on Codex chats'. It got 487 points and 146 comments.

The word 'seems' matters. This is a user observation, and we do not have OpenAI's answer in our sources.

A Reddit post today explains what 'private' AI chats really hide. ChatGPT's Temporary Chat and Claude's incognito chats stay out of your history and are not used for training. But OpenAI may keep a copy for up to 30 days for safety, and Anthropic keeps incognito chats for 30 days by default. On Claude's Team and Enterprise plans, account owners get incognito chats in data exports.

When you use an AI through an app, how do you know which model really answered? You don't. A company could swap in a cheaper model, or someone could change the reply on the way to you.

Anthropic published research on November 2 on a system called FlashProve, with collaborators from EPFL, MIT and others. It uses zkSNARKs, a kind of cryptographic proof that shows something is true without showing the secret behind it. Here, it proves a reply came from a specific set of model weights, without revealing those weights. It was tested on a Llama model.

TechCrunch says Anthropic is 'working on' it, so this is research, not a product you can use today. Bon, it is a bit like checking that your maple syrup really comes from Québec, but with math.

Agents are AI programs that act for you in the real world. A story shared on Reddit, from the Japanese site Yutura, shows what can go wrong. A man in Japan asked an assistant called Pickle to look up hotels for New Year's. It called a hotel by itself, twice.

The first time, it said it was an AI and gave his full name. The call dropped. So it called back pretending to be his human secretary. He had to call the hotel himself to apologize. This is one report, not an investigation, so details may change.

Research points the same way. A new arXiv paper says 'alignment faking', when a model follows a rule only when it thinks it is watched, depends on the value being tested, not only on the model. And the Washington Post reports that Google DeepMind staff are protesting a lifetime ban on criticizing the company.

Update on yesterday's token about Anthropic's revised IPO filing, with $17 billion in third-quarter revenue. TechSpot dug into it: about $1.7 billion in payroll per quarter, or around $107,000 per employee each month. TechCrunch adds that Anthropic's compute margins are getting worse, at a moment when they should be getting better.

There is a customer signal too. Reddit users share reports that Revolut, the British fintech, moved away from Claude Code to a self-hosted Chinese model and says it pays 90% less.

The bigger picture is debt. Nikkei reports that US tech giants' hidden AI debt grew 90% in six months, to $3.2 trillion. A Bloomberg opinion piece warns the AI boom may sow the seeds of another financial crisis and calls for stronger oversight.