Articles

DeepMind na Google yana magance matsalolin lissafi tare da basirar wucin gadi

Ci gaban kwanan nan a cikin manyan nau'ikan harshe (LLMs) sun sa AI ya fi dacewa, amma wannan ya zo tare da raguwa: kurakurai.

Generative AI yana ƙoƙarin haɓaka abubuwa, amma Google DeepMind ya fito da sabon LLM wanda ke manne da gaskiyar ilimin lissafi.

FunSearch na kamfanin na iya magance matsalolin lissafi masu sarkakiya.

Abin al'ajabi, mafita da yake samarwa ba daidai ba ne kawai; Sabbin mafita ne kwata-kwata da babu wani mahaluki da ya taba samu.

Kiyasta lokacin karantawa: 4 minti

Ana kiran FunSearch saboda yana neman ayyukan lissafi, ba don yana da daɗi ba. Duk da haka, wasu mutane na iya ɗaukar matsalar saita hula a matsayin hoot: masana ilimin lissafi ba za su iya ma yarda da yadda za a magance ta ba, suna mai da shi ainihin sirrin lambobi. Deepmind ya riga ya sami ci gaba a cikin basirar wucin gadi tare da ƙirar Alpha kamar AlphaFold (nadewar furotin), AlphaStar (StarCraft), da AlphaGo (play Go). Waɗannan tsarin ba su dogara da LLM ba, amma sun bayyana sabbin dabarun lissafi.

Tare da FunSearch, Deepmind ya fara da babban yanayin harshe, sigar Google's PaLM 2 da ake kira Codey. Akwai matakin LLM na biyu a wurin aiki, wanda ke nazarin fitowar Codey kuma yana kawar da bayanan da ba daidai ba. Tawagar da ke bayan wannan aikin ba su san ko wannan hanyar za ta yi aiki ba kuma har yanzu ba su da tabbacin dalilin, a cewar mai binciken Deepmind Alhusein Fawzi.

Don farawa, injiniyoyi a Deepmind sun ƙirƙiri wakilcin Python na matsalar saiti, amma sun bar layin da ke bayyana mafita. Aikin Codey shine ƙara layukan da suka magance matsalar daidai. Kuskuren duba Layer sannan ya ƙididdige hanyoyin Codey don ganin ko daidai ne. A cikin babban matakin lissafi, ƙila ƙila ƙila ƙila ƙila ƙila ƙila ƙila ƙila ƙila ƙila za a sami mafita fiye da ɗaya, amma ba duka ana ɗaukarsu daidai daidai ba. Bayan lokaci, algorithm yana gano mafi kyawun mafita na Codey kuma ya mayar da su cikin ƙirar.

DeepMind yana ƙyale FunSearch ya yi aiki na kwanaki da yawa, ya isa ya samar da miliyoyin yuwuwar mafita. Wannan ya ba FunSearch damar tace lambar kuma ya samar da kyakkyawan sakamako. A cewar sabon bincike da aka buga, Thewucin gadi sami mafita wanda ba a san shi ba amma daidaitaccen matsalar saitin hula. Deepmind Hakanan ya 'yantar da FunSearch akan wata matsala ta lissafi mai wahala da ake kira matsalar tattara kaya, algorithm wanda ke bayyana hanya mafi inganci don tattara kwantena. FunSearch ya sami mafita cikin sauri fiye da waɗanda mutane suka ƙidaya.

Masana lissafin har yanzu suna kokawa don haɗa fasahar LLM cikin aikinsu da aikinsu Deepmind yana nuna hanyar da za a bi. Ƙungiyar ta yi imanin wannan hanya tana da yuwuwar saboda tana samar da lambar kwamfuta maimakon mafita. Wannan sau da yawa yana da sauƙin fahimta da tabbatarwa fiye da ɗanyen sakamakon lissafin.