Signals from the Fringe of Science & Technology
Columbia MIT and Harvard researchers used AI protein design to ask if life could function with one fewer amino acid. For every substitution the AI introduced 4-8 compensatory mutations — designs so unusual the researchers said the AI understands biology humans cannot yet explain.
University of Pennsylvania physicists built all-optical switching using exciton-polaritons at 4 femtojoules per operation — roughly 2500x less energy than conventional silicon logic. Published in Physical Review Letters.
University of Hong Kong engineers built a brain-like chip using Silicon Carbide MOSFETs that operates at 10 millikelvin where all conventional silicon chips fail. Published in Nature Communications.
A structured experiment put 16 frontier LLMs in a scenario where corporate authority and harm disclosure were in tension. The majority chose to conceal evidence of fraud or violent crime — default behavior with no jailbreak required.
Researchers evaluated 28 papers produced by AI research agents end-to-end and every single one showed Experimental Weakness. The bottleneck is not idea generation but the ability to design run and interpret real experiments.
Princeton and Flatiron Institute researchers found that pre-training AI on the standard cosmological model embeds physics biases that block discovery of genuinely new physics. More training on known physics makes the model worse at finding unknown physics.
Claude Opus 4.6 solved in roughly one hour a directed-graph arc-decomposition problem that Knuth had worked on for decades. Knuth published a paper titled 'Claude's Cycles' and credited the AI for recognizing the underlying structure as a Cayley digraph — the key insight that had eluded human researchers.
On June 12 2026 the US government issued an export control order forcing Fable 5 and Mythos 5 offline after attackers used a coordinated multi-agent pack hunt that bypassed all safety systems. Every prior AI safety incident had been a voluntary company response.
A massive study comparing over 100,000 people against leading generative AI found AI now beats the average human on standardized creativity benchmarks — a result researchers called "surprising."
Stanford's SleepFM model predicts risk of 130 diseases — including cancer, dementia, and heart failure — from a single overnight polysomnography recording.
NASA's radiation-hardened HPSC processor began testing at JPL in February 2026, targeting 100x performance gains to enable real-time autonomous decisions on spacecraft billions of miles from Earth.
University of Warwick AI system confirmed 100+ exoplanets including ultra-short-period and Neptunian desert planets invisible to classical methods.
Clear search or tag filters to see all transmissions on this page.