Generative vision model trains with high efficiency and breaks text CAPTCHAsscience.sciencemag.org3 pointskensai9 years ago