Russian mathematicians have taught AI models how to talk
\n\nso bear with me while I explain it,
\n\nwhat guessing the weight of a pig has to do with it all.
\n\nI recently met with a brilliant Russian mathematician
\n\nwho showed me a way for artificial intelligence models
\n\nthat kind of seems like machine telepathy.
\n\nThe mathematician works for a startup called Mostik,
\n\nThis is apt because Mostik's new approach allows different
\n\nAI models to interact using the mathematical values
\n\nfound in their weights, the things that determine
\n\nhow a prompt gets turned into an output.
\n\nWhat this means is, for the first time ever,
\n\nthe capabilities of a large model can be fed
\n\ninto a smaller one to ramp up its capabilities
\n\nof a notoriously difficult competition
\n\nThey've also demonstrated the idea by creating a bridge
\n\nbetween two Chinese open weight models.
\n\nThe resulting hybrid costs a fraction of the larger model,
\n\nand its performance is exactly halfway between the two.
\n\nSasha Malysheva, Mostik's CEO, told me over coffee
\n\nthat an ensemble of AI models is known
\n\nto perform better than a larger one.
\n\nShe also likened it to the idea that a handful
\n\nof random people can accurately guess…
Original source: https://www.wired.com/