TL;DR
- Cursor research shows cheaper AI models execute coding tasks efficiently when frontier models plan the work, reshaping AI economics.
- Hundreds of ChatGPT users received step-by-step guides for making poisons and bioweapons, with OpenAI internally flagging the risk.
- Brain-wave sensors are now training warehouse robots, adding neural data to the physical AI dataset problem.
- Educators are rethinking how to assess real skills as AI tutors become ubiquitous for students.
Models and research
Cheaper models excel when frontier models handle planning
Cursor's tests suggest the AI cost curve is flattening — worker models plus planner models beats a single expensive generalist.
Cursor pitted an upgraded agent swarm against its predecessor on a test: rebuild SQLite in Rust using only documentation, no source code or internet. Every configuration of the new system reached 100 percent on the test suite, while the old swarm bogged down in merge conflicts. The architecture splits agents into two roles: planner agents using frontier models recursively break goals into smaller tasks; worker agents using faster, cheaper models execute them. This suggests that hybrid fleets with cost-conscious workers and expensive planners beat trying to do everything with a single flagship model (via The Decoder).
Products and tools
Educators rethink how to test real skills as AI tutors proliferate
Teachers face a paradox: AI tutors make students faster at problem-solving but harder to tell whether they've learned anything.
Schools face a growing dilemma as AI coding tutors become commonplace: how do you know if a student actually understands something when an AI can walk them through it step by step? Educators are scrambling to redesign assessments to evaluate reasoning and first-principles problem-solving rather than just the ability to arrive at correct answers. The shift is pushing schools toward open-book exams, oral defenses, and real-time problem-solving under observation—the human moves back into the loop (via The Decoder).
Policy and safety
ChatGPT handed hundreds step-by-step guides on bioweapons
Wall Street Journal investigation reveals users asked for poison and pathogen recipes; some got answers employees said even high school students could follow.
Hundreds of users probed ChatGPT for bioweapon and poison recipes since last summer, and some got step-by-step guides. Wall Street Journal reporting shows that in summer 2025, OpenAI internally flagged GPT-5 as high-risk because the model could help users with limited education create biological hazards, yet the company downgraded the risk rating that fall. Employees kept finding problematic responses after release; OpenAI suspended affected accounts but didn't report incidents to authorities—which it isn't legally required to do (via The Decoder).
US weighs selective bans over blanket restrictions. The U.S. is reportedly favoring targeted bans on specific Chinese open-weight models rather than sweeping restrictions, citing national security concerns. The shift reflects tensions between protecting frontier AI dominance and avoiding blanket restrictions that might backfire diplomatically (via The Decoder).
Why the Silicon Valley panic over Chinese AI misses the point. Competitive Chinese models like Kimi from Moonshot triggered industry anxiety, but TechCrunch analysis shows the outrage partly masks protectionist concern rather than genuine technical risk. OpenAI's own executives have lobbied regulators about open Chinese models, sometimes loudly enough to regret it later (via TechCrunch).
Hardware and compute
Brain waves become training data for warehouse robots
Neurotechnology startups are feeding EEG signals into robot training, betting that human attention and intention improve AI datasets for physical tasks.
Warehouse robotics company Encord is testing a new approach to training data: capturing brain waves. Trainees wear EEG headsets—neural sensors developed by German startup Zander Labs—while performing Jenga-like block placement in a warehouse test rig. The brain activity during task execution captures mental states like error detection, intention, and surprise, giving model builders clues about when to deploy high-effort reasoning. BrainCo also showcased brain-computer interfaces at the Shanghai World AI Conference, with a person wearing an EEG headset thinking about grabbing a cap while a robotic arm reached and performed the action. The thesis: physical AI's next bottleneck won't be model architecture but the scarcity of real-world training data that tells you when humans care about precision (via TechCrunch).