Artificial Intelligence•MIT Students Built a Working Jet Engine With AI, Then Found Where AI BreaksMIT’s JARVIS Challenge asked seven student teams to design and hot-fire small gas-turbine engines with frontier AI copilots. A winning engine generated net thrust, but hallucinations, weak physical understanding and manufacturing delays showed that expert judgment—not the model—remained decisive.
Artificial Intelligence•KAIST’s Temporary ‘Jailbreak’ Makes Personalized AI Fine-Tuning SaferKAIST researchers introduced Buffer-and-Reinforce, a two-stage fine-tuning method that temporarily isolates a language model while it learns private or company data, then strengthens its safeguards. Tests reduced harmful-response rates without materially sacrificing customization, but the evidence remains limited to controlled model evaluations.
Artificial Intelligence•AI Helps Break an 80-Year-Old Erdős Mathematics BarrierA general-purpose reasoning model produced a new construction for Paul Erdős’s planar unit-distance problem. It does not solve the entire question, but overturns a long-standing assumption and demonstrates a new form of human–AI research collaboration.
