Humanity's Last Exam: The Test AI Keeps Failing
2,500 questions no AI can Google. GPT-4o scored 2.7%, humans hit 90%. Inside the hardest AI benchmark and its 30% error rate.
2,500 questions no AI can Google. GPT-4o scored 2.7%, humans hit 90%. Inside the hardest AI benchmark and its 30% error rate.
Gut bacteria from sleep-deprived mice triggered Alzheimer's-like tau damage in healthy brains. Scientists traced the full molecular chain.
A 20-year study of 70,000+ blood samples shows bicarbonate rising in lockstep with atmospheric CO₂ — set to breach safe limits by 2076.
Engineers gave Clostridium sporogenes quorum sensing — bacteria find tumors, wait for backup, then destroy cancer from within.
Experienced devs write code 19% slower with AI. Yet a startup with zero handwritten code sold for $80M. How can both be true?
Genome study of 600K people found 254 genes shaping personality. A 6th trait beyond Big Five predicts mortality. Seven 2024-2025 studies reviewed.
AI-curated database of 67,000 magnetic materials reveals 25 high-temperature alternatives to rare-earth magnets for EVs.
MIT-designed accelerator transmutes long-lived nuclear waste into safe isotopes in 300 years instead of 100,000 — while generating electricity.
Why muscle knots form, how they harm your health, and how to treat them — featuring the molecular pathway discovered in 2024.
Get notified about new articles