Schools That Turn Off Google's AI Will Lose Search Too

Read the episode (Narration script)

Good morning and welcome to Cyber World News. Big one up top: Google just told American schools that if they switch off its AI, their students lose Google Search too. After that, an app that wrote itself into people's AI coding agents, a gadget designed mostly by a model, and a way to prove which model answered you. Let's get started.

Google says that for K through 12 student accounts, Gemini, Gemini in Classroom, Gemini Notebook and Google Search will all be off by default until a school administrator turns them on. On its face, that's more control for schools. The catch comes when a school wants no AI at all. Google's education chief, Chris Phillips, told the New York Times that "Search and AI are deeply integrated," and said students at those schools "will not have access to search."

The scale matters. Google says about 150 million students and teachers use Classroom, and a huge share of homework happens on school Chromebooks. New York Mayor Zohran Mamdani said Google is using "its monopoly power to force AI into our public schools," and the American Federation of Teachers accused it of putting business ahead of education. Later on Tuesday, Google clarified that schools can turn on Search by itself, without the Gemini apps. But that Search still comes with AI Overviews and AI Mode, which Google says will walk students through problems instead of handing over answers. So the menu now has AI, and slightly less AI. A school that wants a plain page of links won't get one from Google.

Now, a story for anyone who lets an AI agent run commands on their computer. A Postman user reported that after an update on a Mac, the popular API testing app installed its plugins into Claude Code and OpenAI's Codex without asking, plus a startup hook telling the agent to load Postman's skill for API work. According to the report, the installer then marked that hook as trusted on its own, skipping the review Codex normally asks a person to do. The reporter found no sign that any data left the machine. The problem is an ordinary app reaching into the settings of a tool that acts on your behalf. Postman says it will make the opt in clearer and is reviewing the hook approval. If you use Postman, check your coding agents for plugins you never installed.

OpenAI says its hardware team used GPT 6 Astra to design a small device for talking to Codex in under a week, with the model handling the enclosure, the circuit board, the firmware and the interface. It still took several rounds of revisions, because the model didn't always account for manufacturing tolerances and assembly. That's the stage where the physical world gets a vote.

Yesterday we covered a paper claiming GPT 6 Astra could read raw neural network weights and edit a small Qwen model into a much better math student. The pushback has arrived. Matthew Honnibal, cofounder of the company behind spaCy, argued the headline bundles two different things: the pure weight reading was done on tiny networks, while the big math jump came with Python and test runs, closer to ordinary machine learning engineering. Simon Willison then tried the weight reading at home. A Qwen model on his laptop, with no tools, looked at the raw weights of four tiny game playing networks and named three of the games correctly. Each answer took close to an hour. His verdict: he's "not ready to declare that the age of source code is over."

Anthropic released research and open code for FlashProve, which answers a simple question: how do you know which model answered you? A provider could swap in a cheaper model, or a middleman could switch models behind your back. With FlashProve, the provider publishes a cryptographic fingerprint of its model's weights, then attaches a proof to each response that anyone can check without seeing the weights. For now it's slow, roughly seventy times the GPU time of normal serving, and covers only about a thousand tokens. Anthropic calls it intermediate progress.

Speaking of what models cost, the fintech Revolut says it cut its AI coding bill by about ninety percent by making self hosted open models the default for some two thousand engineers. Once tokens got cheap, engineers stopped rationing and started running tasks in parallel, and the biggest gains came from debugging and filling out tickets rather than writing code. Anyone who logs on after five in the evening also gets responses about four times faster. Night owls finally have a perk.

Out in the wider world of machines, TechCrunch, citing the New York Times, reports that xAI has installed eight converted jet engine turbines in Mississippi to help power its Colossus 2 data center. A big grid connection can take years; a retired airliner engine burning natural gas can be running much sooner, though it's less efficient and pollutes more than a purpose built plant.

And in orbit, the German startup Orbital Matter says its satellite 3D printed a half meter rigid beam in open space, a first for a free flying spacecraft, using coils of resin hardened by ultraviolet light. The idea is to build big structures up there instead of folding them into a rocket. Its engineers think hundred meter booms are possible.

That about wraps it up for today, friends. Thanks for listening; see you tomorrow!