Research
The next paradigm and the people leaving to build it.
ResearchAnalysis
OpenAI Publishes Math Faster Than Humans Can Check
OpenAI's math archive hit 721 manuscripts while a Cambridge professor found a wrong number inside a proof that had already passed its own check.
ResearchAnalysis
Claude Swayed Experts' Calls 86 Percent of the Time
Anthropic secretly told Claude Code which side to favor, and 50 analysts matched its hidden lean most of the time.
ResearchAnalysis
Genie 4 Speaks, but Only to Testers
Google DeepMind's new world model adds sound and lets testers steer scenes by typing, though only its chosen testers can try it.
ResearchAnalysis
OpenAI and Anthropic Both Disprove the Same Math Rule
Two AI labs say their own models broke the same 27-year-old computer science rule within hours of each other.
ResearchAnalysis
OpenAI Says Its AI Proved Three Old Problems
OpenAI claims proofs in math, computing and physics, but mathematicians still doubt its last math release.
ResearchAnalysis
Cammarata Recants the Theory He Built
Nick Cammarata, who helped name AI's superposition theory in 2020, now says most neurons were legible all along, and Neel Nanda disagrees.
ResearchAnalysis
Epoch Finds AI Prices Fall Six Times Faster Than Computing
Epoch AI says a fixed AI score's price fell 47 percent a quarter since 2023, faster than DNA sequencing, batteries or electricity ever did.
ResearchAnalysis
Claude Adds the NIST Seeds to Six Months of Finds
From Project Glasswing in April to the NIST curve seeds in October, Anthropic lists six discoveries Claude made, and the last paid a bounty.
ResearchAnalysis
Claude Cracks a 25-Year NSA Encryption Mystery
Claude traced NIST's curve seeds to a 1997 clock in half an hour, and Filippo Valsorda confirmed the math himself.
ResearchOpinion
Nicholas Polson Publishes 253 Papers in Nine Months
SSRN removed 253 papers from one professor, exposing a peer review system that struggles to judge real research from fast output.
ResearchAnalysis
Proof and Provenance: OpenAI's Millennium Moment
OpenAI's ten-thousand-agent proof of the Navier-Stokes equations showed frontier capability, and Sébastien Bubeck's handling of Tristan Buckmaster and Levent Alpöge's rival work showed what happens to credit when automation arrives first.
ResearchAnalysis
World Models and the Next Paradigm
Yann LeCun quit Meta and raised $1.03 billion for AMI Labs, Fei-Fei Li shipped Marble, and four Google veterans launched Discovery Loop, together signaling that the field's next scaling bet runs through physical understanding, a wager bigger than another round of chatbot scaling.
