Scientists Discover a Mathematical Reason AI Training Can Be Unpredictable

Researchers at the University of Sydney have identified a mathematical phenomenon that may explain why artificial intelligence systems develop different behaviors despite nearly identical training conditions. Scientists trained approximately 500,000 neural networks and discovered that their learning outcomes can follow intricate fractal patterns called riddled basins. These structures contain interwoven regions in which extremely small changes to initial parameters produce substantially different results. In one image-recognition network, changing a single bit of its starting configuration altered the final internal configuration in approximately 48% of cases. Increasing numerical precision by a factor of 10 billion did not meaningfully improve reproducibility. Published in Nature Machine Intelligence and reported November 5, the findings suggest that some AI unpredictability arises from the mathematics of learning rather than hardware errors or external randomness. (Phys.org)

Researchers Identify a Hidden Weakness That Could Make Cancer Drugs More Effective

Scientists at NYU Langone Health have discovered that certain locations within cancer-cell DNA are exceptionally vulnerable to damage, potentially explaining why some tumors respond strongly to PARP inhibitors. Using advanced molecular imaging and CRISPR screening, researchers identified a protein called ESCO2 that helps protect chromosomes during DNA replication. Reducing ESCO2 caused unusual DNA structures, known as G-quadruplexes, to accumulate near centromeres, regions important for chromosome separation. These structures interfered with DNA copying and made cancer cells more susceptible to PARP-inhibiting drugs. The effect was observed in experimental models involving breast, lung, ovarian and colorectal cancer cells. Published in Nature Cell Biology and reported November 5, the findings suggest that the location of DNA damage may matter as much as its quantity. Clinical applications remain untested. (Medical Xpress)

Cash Payments to Pregnant Mothers Linked to Dramatic Drop in Newborn Deaths

A program providing direct financial assistance to pregnant women and families with infants has been associated with a substantial decline in newborn deaths in Flint, Michigan. Researchers examining statewide birth and death records found that neonatal mortality fell from 10.2 to 5.1 deaths per 1,000 live births following the introduction of Rx Kids in 2024. Deaths during the first seven days of life declined by 74%. The program provides $1,500 during pregnancy and $500 monthly during infancy, helping families cover housing, food and essential supplies. Published in JAMA Pediatrics and reported November 3, the findings complement earlier evidence of reduced premature births and low birth weights. Although the results are encouraging, the observational study cannot establish that cash payments alone caused the improvements. (Medical Xpress)

Ancient Coins Suggest a British King Who Influenced Rome’s Invasion Was a Woman

An archaeologist at Bournemouth University has challenged the longstanding assumption that Verica, an influential Iron Age British ruler, was male. Verica governed the Atrebates in southern Britain before being expelled around AD 40 and seeking assistance from the Roman emperor Claudius. The ruler’s exile became an important justification for Rome’s invasion of Britain in AD 43. Archaeologist Miles Russell argues that Verica’s name, the use of the Latin title rex, and depictions of a female horse-riding warrior on associated coins support identifying the ruler as a woman. The interpretation, discussed in Current Archaeology and reported November 4, could significantly alter historical accounts of female political power in pre-Roman Britain. However, the argument is an archaeological reinterpretation rather than definitive proof, and its acceptance by other historians remains uncertain. (Phys.org)

Underwater Neutrino Telescope Puts Einstein’s Relativity to a New Test

An international team of physicists has used a neutrino telescope deep beneath the Mediterranean Sea to test one of the fundamental principles of Einstein’s theory of relativity. Researchers from the KM3NeT collaboration analyzed 5,828 atmospheric neutrino candidates recorded by the ORCA detector, searching for evidence that the laws of physics change with a particle’s speed or direction. This principle, known as Lorentz invariance, underpins modern physics, although some quantum-gravity theories predict possible violations. The scientists found no such violations but established new limits on deviations involving muon and tau neutrinos. Published in the Journal of Cosmology and Astroparticle Physics and reported November 4, the findings demonstrate the sensitivity of underwater neutrino observatories. Future observations with larger detectors could reveal subtler effects or further constrain alternative theories of spacetime. (Phys.org)