Anthropic open-sources FlashProve, its provable inference prototype for MoE models

Anth­ro­pic’s Fron­tier Red Team, with co­lla­bo­ra­tors at EPFL, MIT and Stark­Wa­re, built Flash­Pro­ve, a pro­toty­pe that pro­du­ces a cry­pto­graphic proof that one spe­ci­fic lan­gua­ge model ge­ne­ra­ted one spe­ci­fic reply. The pro­ver, two ve­ri­fiers, the for­mal proofs behind them and a pro­to­col spe­ci­fi­ca­tion are on GitHub under the Apa­che 2.0 li­cen­se. Sour­ce: Anth­ro­pic When you send a prompt to a hos­ted API, you can’t con­firm which model ans­we­red it. A pro­vi­der could route the re­quest to … More → The post Anth­ro­pic open-sour­ces Flash­Pro­ve, its pro­va­ble in­fe­ren­ce pro­toty­pe for MoE mo­dels ap­pea­red first on Help Net Se­cu­rity .