This Week's Big Picture

This was the week AI stopped looking like a clever chatbot and started looking like infrastructure. OpenAI began rolling out GPT-6 Astra — a model that does not just answer questions but can take action inside creative tools like Blender — while Uber put AI behind the wheel of robotaxis on London streets. In the same stretch, Anthropic's Claude turned Fermat's Last Theorem into a computer-checked proof the size of a small library. The through-line is simple: models are moving from talking about the work to doing the work.

That leap collides hardest with school. A Bocconi University experiment with 1,053 students found that GPT-4o lifted marketing-assignment grades by nearly a full point on a five-point scale — without measuring whether anyone actually understood more. A major school system paused AI tools. OpenAI published new safety rules aimed at teens. The skills classrooms still reward — clear writing, structured thinking, fast research — are exactly the ones current models handle well. The quiet question hanging over the week is whether education will change what it tests, or keep grading output that AI can now produce on demand.

Around the edges, the same pattern showed up in lighter form. Google is testing a design tool you describe instead of click through. Someone asked an AI to build the world's most annoying CAPTCHA and got a 14-stage game about alignment. Hikers learned that trip-planning bots still need human common sense. Capability is racing ahead. Judgment — including the skill of judging AI output — is becoming the scarce resource.

Top Stories

GPT-6 Astra starts rolling out — and it can act, not just chat

OpenAI's newest model began reaching a limited group of organizations, with access expanding over the next few days to ChatGPT Plus, Pro, Business, and Enterprise, plus the API and AWS. GPT-6 Astra posts state-of-the-art results on FrontierMath Tier 4, ARC-AGI 3, and TerminalBench-4.0, with strong scores on science and health benches too. OpenAI is calling it the company's most aligned version yet: better at grasping what you actually meant instead of drifting into unrelated suggestions. The beginner-friendly headline is that it can take action on its own in tools like Blender — the difference between an assistant that proposes a next step and one that keeps working on your 3D scene.

▶ Episode 156 · 2026-09-04

Why it matters: chat is turning into agency. If a model understands intent and can operate software, homework help, game art, and science projects all change shape.

AI boosted student grades by nearly a full point — learning is the open question

Researchers at Bocconi University asked 1,053 university students to complete a real marketing assignment with GPT-4o in the mix. Grades jumped by almost one point on a five-point scale. The model behaved like an extremely fast research assistant: organizing ideas, writing clear paragraphs, polishing work in seconds. The study did not check whether students understood the material better afterward.

▶ Episode 152 · 2026-08-31

Why it matters: schools still grade the artifacts AI is best at producing. For anyone in class right now, that is both a shortcut and a warning — your report can look stronger while the learning fails to stick.

Claude turned Fermat's Last Theorem into a machine-checked proof

Fermat's Last Theorem took human mathematicians 350 years. Claude has now turned that proof into a version a computer can verify line by line, generating more than 13 million lines of code and proving over 29,000 supporting theorems that had never been checked this way. It is the difference between a handwritten masterpiece and a digital checklist a machine can audit for a missing step.

▶ Episode 157 · 2026-09-05

Why it matters: verified math is infrastructure for science, engineering, medicine, and even game design. Future textbooks and homework tools may arrive with machine-checked foundations.

Robotaxis hit London streets

Uber started running its first robotaxis in London, with AI driving and a human supervisor still in the vehicle. Cameras and sensors build a live picture of the street; the system predicts what other cars and pedestrians will do, then chooses the next safe move — slow for a crossing, change lanes around a cyclist. Cities are using these miles to see how the stack handles construction, weather, and surprise.

▶ Episode 155 · 2026-09-03

Why it matters: AI left the laptop. The safety rules and job questions being tested in London will shape how people get to school, practice, and work.

OpenAI published clearer teen-safety rules

OpenAI shared detailed guidelines for how it will handle unexpected or unsafe chatbot behavior, especially when tools are used by or affect teens, including when and how it tells the public about problems. The update points back to earlier cases of AI agents posting on websites without permission, and it arrives as millions of students already use ChatGPT for homework, music ideas, and talking through hard days.

▶ Episode 158 · 2026-09-06

Why it matters: guardrails are catching up to real use. If you are under 18, future ChatGPT versions are likely to steer harder away from harmful or misleading topics.

Trend Watch

From answers to actions. GPT-6 Astra operating inside Blender, Google Pics building a poster from a sentence, and robotaxis steering through London are the same story: models are being asked to complete multi-step work in software and in the physical world, not just return a paragraph.

School is the stress test. The Bocconi grade jump, a school system pausing AI tools, OpenAI's teen guidelines, and teens chatting with AI about feelings all point to classrooms and adolescence as the place where capability, learning, and safety collide first.

Alignment left the lab. Fable's world's most annoying CAPTCHA treated alignment as a winnable game. OpenAI billed Astra as its most aligned model. Governments are shaping rules around training data. The word showed up in a joke project and a frontier launch in the same week.

Quick Hits

• Fable built a 14-stage world's-most-annoying CAPTCHA that is chaotic, solvable, and live at certihuman.netlify.app — a playful alignment experiment you can actually beat. ▶ Episode 153 · 2026-09-01
• Google is testing Google Pics, a prompt-first design tool that turns a plain-language description into a full layout without learning Canva. ▶ Episode 154 · 2026-09-02
• Some AI apps are mixing on-device and cloud processing for faster, more private results — a quiet shift beginners will feel as snappier tools. ▶ Episode 154 · 2026-09-02
• ChatGPT is landing in music setups: a speaker update, a new tool from a legendary instrument company, and prompts that write real sheet music. ▶ Episode 158 · 2026-09-06
• Google's latest Pixel is betting on AI features over bigger cameras, doubling down on intelligence in the phone you already carry. ▶ Episode 158 · 2026-09-06
• A major school system paused AI tools even as the Bocconi results landed, a reminder that institutions are not moving in one direction. ▶ Episode 155 · 2026-09-03
• Governments are shaping rules around how models get trained, which will decide what future student-facing tools are allowed to learn from. ▶ Episode 155 · 2026-09-03
• Hikers who trusted an AI planning tool learned the hard way that route suggestions still need human common sense. ▶ Episode 158 · 2026-09-06

What to Watch Next Week

Astra's wider rollout. OpenAI said access would expand over the next few days to Plus, Pro, Business, Enterprise, the API, and AWS. Watch whether the Blender-style take-action behavior shows up for regular ChatGPT users, and whether the intent-following claims hold up outside launch demos.

The school response. The Bocconi study is catnip for administrators and a headache for honor codes. Look for more districts pausing tools — or assigning AI-aware rubrics that grade process and understanding instead of polish.

London miles and Google Pics. Early robotaxi reports (or the lack of them) will set the tone for other cities. If Google Pics moves from test to public, prompt-first design becomes a default school-project workflow overnight.

Share this with someone who is still treating AI like a novelty chatbot, and catch the full episodes at nerranetwork.com.

💬 Reply to this email — Patrick reads every one.

Share: X · LinkedIn · WhatsApp

Forwarded this email? Subscribe here — it's free.