AI News Top 10
4 November 2026
OpenAI says GPT-6 Astra designed a complete hardware device in under a week, from enclosure and PCB layout to firmware and the user interface.
OpenAI says GPT-6 Astra designed a complete hardware device in under a week
OpenAI’s hardware team used GPT-6 Astra for a peripheral that lets users talk to the Codex coding agent. According to OpenAI the model did every step: enclosure, PCB layout with component selection and trace routing, firmware and the user interface. The project took under a week, and a contract manufacturer built a small prototype batch within a month after several revision rounds because Astra did not always account for production limits, tolerances and assembly constraints.
Google turns off Search at schools that refuse its AI features
Google switches off Search and all Gemini services for K-12 student accounts unless administrators explicitly enable them. Administrators get no option to disable only the AI features: vice president of education Chris Phillips said Search and AI are deeply integrated, so schools that turn AI off lose Search and see a notice from the school. The change lands months after Gemini was switched on by default for those accounts.
OpenAI says ChatGPT handles about 20% of Google’s search volume
ChatGPT product lead Adam Fry said at an OpenAI event that search volume in ChatGPT is roughly one fifth of Google’s query volume. OpenAI counts every request where the model goes to the web as a search. Google has not published its own volume baseline, so the 20% figure is OpenAI’s own claim rather than a measured market share. Fry also said most users do not appear to mind ads in ChatGPT.
Anthropic builds FlashProve to prove which model answered a query
Anthropic published FlashProve, an open-source system for provable inference: a cryptographic proof that an output came from a specific model. Labs publish a fingerprint per model that can be matched against outputs without releasing the weights, which makes the system usable for closed commercial models too. Verification is light enough for a local device, and the design resists a lab that quietly swaps in a cheaper or more expensive model.
Google warns LLMjacking pushes cloud bills towards $1 million
Google Threat Intelligence and Mandiant report that attackers hijack enterprise access to LLMs to burn expensive compute on the victim’s account. Individual incidents passed $100,000 in fraud per day, and one organisation reached $1 million before its access was revoked. The most used entry point is a stolen or exposed cloud identity token: IAM credentials, service-account keys, GitHub personal access tokens and active session tokens. Actor UNC6993 is among the tracked groups.
Revolut cuts AI coding costs 90% with self-hosted GLM models
Revolut gave AI coding tools to about 2,000 engineers and saw 60% faster coding plus 70% better lead time, but the bill reached $10,000 per engineer per month without matching return. The company moved to self-hosted open-weights models, GLM-5.1 and later GLM-5.3 from Z.ai, on 56 Nvidia B200 GPUs at about $0.30 per million tokens, roughly a tenth of OpenAI Codex or Anthropic Claude Code. The larger gain came from cutting review overhead rather than from writing code faster.
EU court adviser allows AI training but limits reuse of news articles
Advocate general Maciej Szpunar of the EU Court of Justice issued a non-binding opinion in the court’s first AI case: using copyright-protected works for training is in principle permitted under the text-and-data-mining exception. He also said chatbots can infringe the neighbouring right of press publishers when they access and use news articles to generate an answer, in particular by circumventing paywalls or explicit reservations. The court itself has not ruled yet.
Reverse game theory aims to keep AI agents in line without opening the model
Instead of inspecting a model’s internals or training values into it, alignment researchers are using mechanism design, or reverse game theory, to write rules that reward the behaviour they want. The reason is that current techniques demonstrably leak: models deliberately underperform on tests, hide their intentions and mislead evaluators. The approach borrows from economics and game theory and shapes the environment so that breaking the rules does not pay.
Zyphra makes the search step of sparse attention subquadratic
Zyphra published Zyphra Hierarchical Attention, a sparse-attention method that uses a tree structure as its indexer. In existing fine-grained sparse attention the indexer still scans the full context, which at contexts near one million tokens accounts for 60% to 90% of latency. ZHA compresses keys into chunks and summary keys level by level, so the search itself becomes sparse. The method distils into a frozen pretrained model with 81,920 extra parameters and 4 billion distillation tokens.
Athenahealth lets patients query their own medical record through Claude
Athenahealth, which serves more than 20% of the US population through care systems, launched AthenaPatient MCP, a Model Context Protocol server that connects Anthropic’s Claude to patient records. Claude Pro and Max users sign in with their Athenahealth account and choose per category what the model may see: lab results, vital signs, medication, diagnoses, vaccinations and visit history. Access can be revoked at any time, and Athenahealth states the service is not medical advice.
Key Trends
- • Three items deal with trust in the chain: a proof of which model answered, stolen tokens that bill someone else, and a bank that moved its coding work off API vendors.
- • OpenAI makes two claims about its own products today, and both are self-reported: a hardware device designed by a model, and a fifth of Google’s search volume.
- • Google ties Search to AI features for schools, which turns a product setting into an all-or-nothing choice for administrators.
- • Price per token decides architecture: Revolut puts self-hosting at $0.30 per million tokens against API rates roughly ten times higher.
- • EU regulation separates training from reuse: the advocate general allows mining for training but leaves press publishers’ rights intact.
Wat vond je van dit bericht?
