I-DeepMind yakwaGoogle ixazulula izinkinga zezibalo ngobuhlakani bokwenziwa

Ukuthuthuka kwakamuva kumamodeli wezilimi ezinkulu (LLMs) kwenze i-AI yavumelana nezimo, kodwa lokhu kuza nobubi: amaphutha.

I-Generative AI ijwayele ukwenza izinto, kodwa i-Google DeepMind iqhamuke ne-LLM entsha enamathela emaqinisweni ezibalo.

I-FunSearch yenkampani ingaxazulula izinkinga zezibalo eziyinkimbinkimbi.

Ngokuyisimangaliso, izixazululo elizikhiqizayo azigcini nje ngokunemba; zingamakhambi amasha ngokuphelele okungekho muntu owake wawathola.

Isikhathi sokufunda esilinganiselwe: 4 imizuzu

I-FunSearch ibizwa kanjalo ngoba isesha imisebenzi yezibalo, hhayi ngoba imnandi. Kodwa-ke, abanye abantu bangase bacabangele inkinga yesethi ye-cap njenge-hoot: izazi zezibalo azikwazi ngisho nokuvumelana ngokuthi zingayixazulula kanjani kangcono, ziyenze ibe imfihlakalo yezinombolo zangempela. Deepmind isivele yenze intuthuko kwezobuhlakani bokwenziwa ngamamodeli ayo e-Alpha afana ne-AlphaFold (ukugoqa amaprotheni), i-AlphaStar (StarCraft), ne-AlphaGo (playing Go). Lezi zinhlelo bezingasekelwe ku-LLM, kodwa zembula imiqondo emisha yezibalo.

Nge-FunSearch, Deepmind iqale ngemodi yolimi enkulu, inguqulo ye-Google PaLM 2 ebizwa nge-Codey. Kukhona izinga le-LLM lesibili emsebenzini, elihlaziya okuphumayo kwe-Codey futhi lisuse ulwazi olungalungile. Ithimba elenza lo msebenzi belingazi ukuthi le ndlela izosebenza yini futhi namanje alikaqiniseki ukuthi kungani, ngokusho komcwaningi Deepmind Alhussein Fawzi.

Ukuqala, onjiniyela ku Deepmind badale ukumelwa kwePython kwenkinga yesethi ye-cap, kodwa bashiya ngaphandle imigqa echaza isisombululo. Umsebenzi kaCodey bekuwukwengeza imigqa eyixazulule ngokunembile inkinga. Iphutha lokuhlola isendlalelo libe selithola amaphuzu ezixazululo ze-Codey ukuze libone ukuthi zinembile yini. Ezibalweni zezinga eliphezulu, izibalo zingase zibe nezixazululo ezingaphezu kwesisodwa, kodwa akuzona zonke ezibhekwa njengezinhle ngokulinganayo. Ngokuhamba kwesikhathi, i-algorithm ikhomba izixazululo ezingcono kakhulu ze-Codey futhi izibuyisele kumodeli.

I-DeepMind ivumela i-FunSearch ukuthi isebenze izinsuku ezimbalwa, isikhathi eside ngokwanele ukukhiqiza izigidi zezixazululo ezingaba khona. Lokhu kuvumele i-FunSearch ukuthi icolise ikhodi futhi ikhiqize imiphumela engcono. Ngokocwaningo olusanda kushicilelwa, Theubuhlakani bokufakelwa ithole isisombululo esingaziwa ngaphambili kodwa esilungile senkinga yesethi ye-cap. Deepmind iphinde yakhulula i-FunSearch kwenye inkinga enzima yezibalo ebizwa ngokuthi inkinga yokupakisha isiqukathi, i-algorithm echaza indlela ephumelela kakhulu yokupakisha iziqukathi. I-FunSearch ithole isixazululo ngokushesha kunalezo ezibalwe abantu.

Izazi zezibalo zisazabalaza ukuhlanganisa ubuchwepheshe be-LLM emsebenzini wabo kanye nomsebenzi we Deepmind ikhombisa indlela okungenzeka ilandelwe. Ithimba likholelwa ukuthi le ndlela inamandla ngoba ikhiqiza ikhodi yekhompyutha kunesixazululo. Lokhu ngokuvamile kulula ukukuqonda nokuqinisekisa kunemiphumela yezibalo eluhlaza.

Ukufundwa Okuhlobene

• Akuseyona inhlanhla nje: ibhola elimnandi liba imodeli yokuhlaziya. UFantalgoritmo wazalwa, lapho umhlaba webhola elimnandi uhlangana nezibalo
• I-Gato, i-DeepMind kanye nomjaho obheke ku-intelligence yokwenziwa ejwayelekile
• Iyini i-Predictive Analytics, futhi kungani ibalulekile ebhizinisini
• IZAKHIWO ZEZINHWEBO EZISIMILE: NGOKUPHATHWA KWAMANDLA KANYE NOBUHLAKANI OBUYIQINISO, UKUSETSHENZISWA KWAMANDLA OMHLABA WONKE KUNGEHLISWA NGO-30% KANYE NOKUKHISHWA KWE-CO37 NGO-2%.
• Ukuqamba kanye nobudlelwano bekhasimende - isipiliyoni kanye nobuhlakani obubambekayo bokudonsa

BlogInnovazione.it