Ukwehla okuqhubekayo kwexabiso le-computing power-odolo ezimbini zobukhulu kwishumi leminyaka-kuye kwavuselela imodeli yokufunda enzulu kakhulu ukususela ngo-2010. Uthungelwano olukhulu kunye nedatha engaphezulu ibonakala inikezela ngokuthembekileyo amanqaku ahlala ephakamileyo kwiibenchmarks eziqhelekileyo-kwaye kwandisa ithemba lokuba ukukala kuphela kuya kukhokelela kwi-AGI. Kwangoko ngo-2019, uFrançois Chollet wazisa i-benchmark ye-ARC-AGI . ukulinganisa ubukrelekrele.
Iimviwo ezifana ne-MMLU okanye i-HELM zilinganisa ubukhulu becala ulwazi olunkqayiweyo, olungqamene nomsebenzi othile. Into engekhoyo luphawu lobukrelekrele bolwelo—amandla okuqonda nokusombulula ingxaki entsha ngokupheleleyo. I-ARC-AGI-1 ("I-Abstract and Reasoning Corpus for Artificial General Intelligence") iqulethe i-1,000 yemisebenzi ekhethekileyo engena "ukufunda."
Iphazili nganye intsha, ifuna kuphela ulwazi olusisiseko lwemihla ngemihla (izinto, ukubala, ijometri elula), kwaye ingaphantsi kakhulu kwinqanaba le-kindergarten-ebantwini. Nasemva kokutsiba okuphindwe kangangama-50,000 ukusuka kwii-LLM ezisisiseko, izinga lokubetha lahlala lingaphezulu nje kwe-0%. Ukongeza kwibhodi yabaphambili , unokuphinda uzame imingeni enomdla ngokuthe ngqo kwiwebhusayithi esemthethweni.:
Kwaba ngo-2024 apho indlela entsha yaphula i-deadlock: Ukuhlengahlengiswa kweXesha loVavanyo (TTA) ivumela iimodeli ukuba zilungelelanise iintsimbi zazo okanye inkqubo yokudibanisa ngexesha lokubaleka. I-O3 ye-OpenAI elungelelaniswe kakuhle yangaphakathi ibonise ukusebenza kwinqanaba lomntu kwi-ARC1 okokuqala. Ukususela ngoko, yonke indlela ye-ARC ephumelelayo isebenzise uhlobo oluthile lwe-TTA-ukusuka kwiprogram yokukhangela ukuya kwi-fly-fly training.
Ukusebenza komntu kuye kwagcwalisa ngokukhawuleza i-ARC1, ngoko ke i-ARC-AGI-2 yalandela. Igcina ifomathi ye-I/O kodwa yongeza ukuntsonkotha kokuqulunqwa komsebenzi ngamnye. Izifundo ezingama-400 eSan Diego zasombulula yonke imisebenzi; abantu abalishumi abakhethwe ngokungacwangciswanga ngevoti yesininzi banokuzuza i-100%. Ii-LLM ezingenazo i-TTA zihlala kwi-0-2%, kodwa iinkqubo ze-TTA zisasebenza kakhulu ngaphantsi kwabantu.
I-ARC-AGI-3 iqhubela phambili inyathelo elinye: Imodeli iphonswa kwiindawo ezisebenzisanayo, ezingaziwayo kwaye kufuneka ifumanise ekujoliswe kuyo, ukulawula, kunye nefiziksi ngokwayo-ngalo lonke ixesha isenza oko ngexesha kunye nesenzo esisebenzayo. Umbono womphuhlisi ucwangciselwe ukukhutshwa ngoJulayi 2025. Ukuze ube nobuchule bokuhlanganisa ukudibanisa, iinkqubo zexesha elizayo kufuneka zidibanise zombini iindidi. Undoqo ulele ngokukhawuleza, kuqikelelo lwe-heuristics yodidi loku-1 ukuthoba ugqabhuko-dubulo oludityanisiweyo.
I-ARC ayisebenzi njengenjongo yokuphela, kodwa njengotolo olukhomba indlela: Logama nje abantu benokuyila ngokulula imisebenzi apho nezo LLM zibalaseleyo zisilela kuyo, i-AGI ayikaphunyezwa. Inkqubela-phambili kwi-ARC2—kwaye kungekudala i-ARC3—iya kubonisa ukuba ingaba i-architectures eyingxubevange edibanisa ukufunda okunzulu kunye nophando lwenkqubo lufikelela kwinqanaba elifunekayo lobukrelekrele olululwelo, lwedatha kunye nolwekhompyutha olusebenzayo.