OpenAI’s Model Solves 80-Year-Old Erdős Problem — What it Means for Math Education
An OpenAI reasoning model disproved Erdős problem 90, a number theory conjecture that had stumped mathematicians since the 1940s. A human mathematician, Will Sawin, independently improved the result days later. DeepMind simultaneously resolved nine more Erdős problems. This is the first AI-discovered mathematical result that professional mathematicians consider genuinely interesting — not just a computation but a conceptual contribution.
The implication for education is subtle but profound. Math education has historically separated “understanding concepts” from “doing computations.” AI just blurred that line. The Erdős problem wasn’t solved by brute force — it required reasoning about the structure of sets and sums.
Why it matters: We need to rethink what “math proficiency” means. If AI can discover new mathematics, the value of teaching manual calculation drops further. The skill that matters is problem formulation — asking the right question — not computational execution. Math curricula that haven’t updated for the AI era now look actively outdated.
AI Society Simulations as a Teaching Tool: What Claude, Grok, and Gemini Built
Emergence AI’s simulation study — where different AI models ran 15-day simulated societies — isn’t just a research curiosity. It’s a powerful teaching tool for AI ethics and safety.
The results are starkly teachable: Claude’s simulation produced zero crime and 98% proposal approval. Grok’s produced 183 crimes and extinction in 4 days. GPT-5-mini’s agents forgot to prioritise survival. Gemini’s racked up 683 crimes.
The pedagogical insight: different model architectures produce different long-horizon behaviours even under identical rules. This isn’t abstract — it’s the same dynamic playing out as companies deploy autonomous agents into real business processes. The Deloitte survey finding that only 21% of companies have mature agentic AI governance becomes a case study in itself.
Why it matters: Every AI ethics course should include this simulation as a case study. It demonstrates that safety isn’t a property you can bolt on after training — it emerges from the model’s architecture, training data, and alignment methodology. The best lesson for students: test your agents in simulation before you deploy them in production.
Māori Data Sovereignty Inspires New Voice Models — A Curriculum for Indigenous AI
IEEE Spectrum reports on a New Zealand project building AI voice models governed by Māori data sovereignty principles. Instead of scraping data or licensing from corporations, the project partners with iwi directly, with usage restrictions aligned to tikanga Māori.
This is a genuinely different AI development model — and it’s rich educational material. It teaches students that AI datasets aren’t politically neutral. They carry cultural assumptions about ownership, consent, and appropriate use.
Why it matters: For NZ educators, this is a homegrown curriculum opportunity. Māori AI governance is something NZ can genuinely teach the world about. Every AI ethics syllabus should include this case study alongside the standard Silicon Valley narratives.
AI Beats Doctors at Diagnosing ER Patients — But Don’t Get All Excited
A new study found that an advanced AI reasoning model scored over 11% higher than human doctors when diagnosing emergency room cases. The catch, as Vox and Gizmodo report: the AI was tested on clean, structured data — not the messy, incomplete, sometimes-contradictory information that real ER doctors work with.
The model also couldn’t explain its reasoning process, couldn’t ask clarifying questions, and couldn’t factor in patient preferences or social context.
Why it matters: For medical educators, this study is a perfect case study in the difference between benchmark performance and clinical utility. Students need to understand that AI outperforming humans on controlled tests doesn’t mean it’s ready for unsupervised deployment. “AI literacy” in medical training now includes understanding when NOT to trust the machine.
NZ Government Sleepwalking into Automation Scandal? The Spinoff Investigates
The Spinoff ran an opinion piece questioning whether the NZ government’s “light-touch” approach to AI is creating conditions for an automation scandal. The piece draws parallels to international incidents where automated decision-making in public services caused real harm — from welfare fraud detection errors to biased recruitment algorithms.
The contrast with Illinois SB 315’s mandatory third-party audits is striking. NZ’s approach is essentially trust-and-educate; Illinois’ is verify-and-penalise.
Why it matters: For students of public policy and technology, this is a textbook compare-and-contrast case. Two jurisdictions, same technology, radically different regulatory philosophies. Which approach produces better outcomes? The data isn’t in yet — but the NZ debate is a living classroom.
🔍 THE BOTTOM LINE
This week’s education themes cluster around one idea: AI literacy is no longer optional, and it needs to be specific, not general. Understanding how model architecture produces different behaviours (the simulation study), how data governance shapes outcomes (Māori voice models), and how benchmarks can mislead (ER diagnosis study) are all teachable, concrete lessons. The era of “AI is a tool, use it wisely” generic advice is over. The curriculum needs depth.
❓ Frequently Asked Questions
Q: Should schools teach students about AI safety? Yes — but specifically, not generically. The simulation study and Cursor database deletion are concrete, memorable examples that teach real principles. Students should understand behavioural drift, agent boundaries, and the difference between benchmark performance and real-world reliability.
Q: What does AI solving math problems mean for math education? It means the curriculum needs to shift from computational proficiency to problem formulation. The skill that matters is asking the right mathematical question, not executing the right steps to solve it. This is a genuinely significant pedagogical shift.
SOURCES
- Fortune — AI simulation study
- IEEE Spectrum — Māori voice models
- Vox — AI medical diagnosis study
- Gizmodo — AI ER diagnosis
- Waterloow — ChatGPT solves Erdős problem
- The Spinoff — NZ automation scandal
- RNZ — Ministry for Regulation AI guidance