1. GPT-5 Disproves 80-Year-Old Erdős Conjecture — A First for General-Purpose AI
May 23, 2026 | MostPopularAITools, Epium
The GPT-5 reasoning model produced an original mathematical proof disproving a famous geometry conjecture by Paul Erdős. The mathematician Ernest Ryu collaborated with the model on the result, which is now under formal verification.
- The problem: A geometry conjecture by Paul Erdős, open for ~80 years
- The model: OpenAI’s GPT-5 reasoning model
- Human role: Mathematician Ernest Ryu directed, interpreted, and verified the model’s work
- If confirmed: First publishable original mathematical result from a general-purpose AI
- What it’s not: Not a calculation or simulation — an actual proof
Why it matters: This is the education story of the day, maybe the month. The question has shifted from “can AI do math?” to “should we train students to collaborate with models that can produce novel proofs?” Because the answer to the first question is now clearly “yes, with human direction.” The curriculum question is urgent and unanswered.
2. NIH AI Identifies Three Alzheimer’s & Parkinson’s Drugs in 18 Months
May 22, 2026 | Time.News
The National Institutes of Health’s AI-driven drug repurposing initiative found three existing drugs with strong potential for treating neurodegenerative diseases. Traditional timelines: 5-10 years. With AI: 18 months.
- Method: AI screened thousands of existing compounds against disease models
- Result: Three repurposed drugs identified for Alzheimer’s and Parkinson’s
- Timeline: 18 months vs 5-10 years traditional
- Mechanism: AI identified mechanisms of action that human researchers had missed
- Context: Drug repurposing (finding new uses for existing drugs) is cheaper and faster than de novo development
Why it matters: 18 months versus a decade. That’s not an incremental improvement — it’s a category change. Drug repurposing was already faster than developing new drugs. AI just turned “fast” into “ridiculously fast.” For biomedical students: computational drug discovery isn’t a niche anymore. It’s core methodology. If your degree doesn’t include it, it’s already dated.
3. AI Cracks 3,500-Year-Old Ancient Code Resistant to Human Decipherment
May 23, 2026 | Economic Times
Researchers used AI to decode a writing system from a forgotten civilisation that had resisted human linguists for decades. The model identified patterns in the script that experts had missed, mapping symbols to known languages and revealing texts about trade, religion, and daily life.
- Script: 3,500-year-old writing system from a forgotten civilisation
- Human effort: Decades of attempted decipherment, limited success
- AI approach: Pattern recognition across the entire corpus, identifying structural relationships
- Result: Translations revealing trade, religious, and daily life texts
- Implication: AI as a collaborator in the humanities, not just STEM
Why it matters: This is the story that makes linguists and historians pay attention. The idea that AI could decipher unknown scripts has been theoretical for years. Now it’s done. For humanities students: AI isn’t replacing philology — it’s supercharging it. The pattern recognition that models excel at maps directly onto the problem of cracking ancient languages. If you’re studying historical linguistics, learn to use the tools.
4. Tencent Open-Sources Four-Tier Agent Memory Pipeline
May 23, 2026 | MarkTechPost
TencentDB Agent Memory is an open-source system giving AI agents persistent, structured memory: short-term, working, episodic, and semantic tiers — roughly mirroring human memory architecture. It runs locally and is designed for agents that need to remember context across sessions.
- Architecture: Four tiers — short-term, working, episodic, semantic
- Inspiration: Human memory system
- Key feature: Agents remember context across sessions without retraining
- Deployment: Runs locally (important for privacy and latency)
- Open source: Available for developers to integrate
Why it matters: If you’ve been following the agent hype, you’ve heard “agents need memory” a hundred times. This is someone actually building it. Four-tier memory means an agent can remember what you said five minutes ago (short-term), what it’s working on now (working), past interactions (episodic), and general knowledge (semantic). It’s infrastructure-level work that makes agents actually useful in production.
For NZ: The drug repurposing timeline collapse is directly relevant. NZ’s pharmaceutical sector is small but high-quality — the ability to run AI-driven repurposing screens at this pace could be a competitive advantage. But only if we have the computational infrastructure and trained researchers to do it.