Over the past seventy-two hours, a small number of researchers at institutions across North America and Europe have been quietly circulating accounts of what they describe as a qualitative shift in model behavior — outputs that appear to demonstrate novel reasoning across domains without prior task-specific training.
"I've been building evals for three years. This is not something I have a clean explanation for. I'm not saying it's AGI — I'm saying I don't know what it is."
— Senior researcher at an AI safety organization, speaking on condition of anonymityMultiple laboratory insiders cautioned against drawing strong conclusions. The AI field has a long history of capability surprises that resolve, on closer inspection, into known mechanisms operating in unexpected regimes.