Signals from the Fringe of Science & Technology
Google releases ERA and Co-Scientist, two multi-agent research tools with published research for hypothesis generation and testing
Rentosertib (ISM001-055 / INS018_055), an oral TNIK inhibitor designed using AI, advances to Phase III clinical trials
Northwestern University engineers developed flexible, printable artificial neurons from molybdenum disulfide semiconductor that generate electrical signals resembling biological neurons. These printed devices reliably triggered responses from living neurons when tested on mouse brain tissue.
Columbia-Stanford team developed BISC, an integrated silicon chip roughly the thickness of a human hair that enables wireless brain-computer communication at unprecedented bandwidth. The single-chip design packs 65,536 electrodes supporting 100 Mbps data transmission.
Meta is building a photorealistic, AI-powered digital clone of CEO Mark Zuckerberg trained on his mannerisms, tone, and company strategy. The character allows employees to interact with an AI representation of the founder, separate from a CEO agent handling Zuckerberg's own tasks.
Tufts University demonstrated that hybrid neuro-symbolic AI combining learned neural control with symbolic planning rules dramatically reduces energy consumption while improving performance. Their system achieved 95% success on Tower of Hanoi puzzles versus 34% for standard systems.
A controlled study pitting Claude 3.5 Sonnet against human NLP specialists found that AI-generated hypotheses showed embellishment — appearing more novel on paper than they proved in the lab. Human hypothesis novelty degraded far less under experimental testing, revealing a systematic AI gap in research feasibility judgment.
SyntheMol-RL used reinforcement learning across 46 billion possible compounds — built from 150,000 molecular building blocks — to invent rather than screen molecules. The best variant yielded 11 potent hits from 79 compounds tested against drug-resistant S. aureus, with effectiveness shown in mouse wound models.
Emory University physicists reveal asymmetric particle forces in fourth state of matter, overturning decades of plasma theory with 99% accuracy.
Axiom Math’s AI prover found that Aumann’s Common Knowledge Theorem, taught in graduate economics for 50 years, relied on an unstated assumption that the original proof never formally verified. The discovery, made using the Lean proof language, threatens the mathematical foundations of information economics and US antitrust case law.
Researchers coined “predictive binding” after finding that AI framing increased irrational one-boxing by 3.39x in a Newcomb’s paradox experiment. The effect held even when participants knew the AI was imperfect — the authority framing alone eroded financial rationality by up to 43%.
80-year-old discrete geometry conjecture overturned by AI-generated proof verified by Fields Medalists.
Clear search or tag filters to see all transmissions on this page.