703
Microsoft’s AI Beat Generalist Doctors on Written Tests. Patient Trials Are Next
It scored 83.2% on complex cases, but the study did not test treatment decisions or patient outcomes.
703
It scored 83.2% on complex cases, but the study did not test treatment decisions or patient outcomes.
792
Google’s AI agents raised an aptamer hit rate to 42.1% across three robotic lab rounds, but the result is research, not a treatment.
931
The bacteriocin has computational and genomic clues behind it, but only laboratory tests can establish whether it works against bacteria and is safe.
878
Anthropic says Claude matched five NIST curve seeds to timestamp-seeded generator output. The finding eases one concern, but leaves other design choices unexplained.
643
Their proposed physics model could support several analysis tasks, but its budget, schedule and eventual performance remain unknown.
525
A Nature study reports stronger forecasts two to six weeks ahead. The gains improve odds of broad conditions, not promises about a town’s weather.
591
The experiment reached 0.39 tokens per second, but left a crucial gap between running AI and proving a self-spreading infection.
693
Epoch AI’s October 7 evaluation put Sol at 15% of a human method’s gains under strict scoring, exposing how selected runs complicate comparisons.
580
Five studies across four countries found gains in select tasks, but no evidence yet of a live public-health warning system.
608
In short online tests, AI users answered fewer follow-up questions correctly and skipped more. The researchers did not measure lasting effects.
527
A 26-researcher paper says inference-only chips could help verify a decade-long pause, but only if rival governments agree on rules and inspections.
695
The manuscripts trace back to about 4,000 problems. OpenAI says formal checking is incomplete, so 722 papers does not mean 722 verified breakthroughs.