OpenAI\u00a0v\u0259\u00a0Google DeepMind\u00a0kimi n\u0259h\u0259ngl\u0259r\u00a0s\u00fcni intellekt\u00a0modell\u0259rini v\u0259 hesablama g\u00fcc\u00fcn\u00fc geni\u015fl\u0259ndirm\u0259y\u0259 \u00fcst\u00fcnl\u00fck verdikl\u0259ri bir vaxtda,\u00a0Thinking Machines\u00a0Lab\u00a0t\u0259dqiqat\u00e7\u0131s\u0131\u00a0Rafael Rafailov\u00a0superintellektin yarad\u0131lmas\u0131 \u00fc\u00e7\u00fcn alternativ yana\u015fma t\u0259klif edib. O bildirib ki, s\u00fcni intellektin g\u0259l\u0259c\u0259yi t\u0259limin h\u0259cmin\u0259 deyil, onun keyfiyy\u0259tin\u0259 ba\u011fl\u0131d\u0131r. Onun s\u00f6zl\u0259rin\u0259 g\u00f6r\u0259, ilk superinsan s\u0259viyy\u0259sind\u0259 s\u00fcni intellekt g\u00fcc\u00fc il\u0259 deyil, \u00f6z t\u0259cr\u00fcb\u0259sind\u0259n \u00f6yr\u0259nm\u0259k v\u0259 uy\u011funla\u015fmaq bacar\u0131\u011f\u0131 il\u0259 f\u0259rql\u0259n\u0259c\u0259k. M\u0259hz bu c\u00fcr sisteml\u0259rin \u00fcz\u0259rind\u0259 OpenAI-nin ke\u00e7mi\u015f texniki direktoru\u00a0Mira Murati\u00a0t\u0259r\u0259find\u0259n yarad\u0131lm\u0131\u015f v\u0259 art\u0131q 2 milyard dollar investisiya c\u0259lb etmi\u015f Thinking Machines Lab startap\u0131n\u0131n i\u015fl\u0259y\u0259c\u0259yi ehtimal olunur. \u201c\u0130lk\u00a0superintellekt\u00a0insan\u0131 \u00f6yr\u0259nm\u0259 qabiliyy\u0259tin\u0259 g\u00f6r\u0259 gerid\u0259 qoyacaq\u201d – dey\u0259 Rafael Rafailov San-Fransiskoda ke\u00e7irilmi\u015f\u00a0TED\u00a0AI konfrans\u0131nda bildirib.<\/p>\n
Onun s\u00f6zl\u0259rin\u0259 g\u00f6r\u0259, bel\u0259 bir s\u00fcni intellekt anlay\u0131\u015flar\u0131 d\u0259rin \u015f\u0259kild\u0259 d\u0259rk ed\u0259, \u00e7evik \u015f\u0259kild\u0259 uy\u011funla\u015fa, \u00f6z f\u0259rziyy\u0259l\u0259rini ir\u0259li s\u00fcr\u0259, t\u0259cr\u00fcb\u0259l\u0259r apara v\u0259 fikirl\u0259rini yoxlamaq \u00fc\u00e7\u00fcn \u0259traf m\u00fchiti istifad\u0259 ed\u0259 bil\u0259c\u0259k. Rafailov, milyardlarla dollar investisiya yat\u0131raraq modell\u0259rin v\u0259 m\u0259lumatlar\u0131n miqyas\u0131n\u0131 art\u0131rma\u011fa \u00e7al\u0131\u015fan OpenAI,\u00a0Anthropic\u00a0v\u0259 Google DeepMind \u015firk\u0259tl\u0259rinin yana\u015fmas\u0131n\u0131 t\u0259nqid edib. Onun fikrinc\u0259, s\u00fcni intellekt sisteml\u0259rin\u0259 \u00e7at\u0131\u015fmayan \u015fey hesablama g\u00fcc\u00fc deyil, \u00f6z t\u0259cr\u00fcb\u0259sind\u0259n \u00f6yr\u0259nm\u0259k qabiliyy\u0259tidir. Problemi izah etm\u0259k \u00fc\u00e7\u00fcn t\u0259dqiqat\u00e7\u0131 s\u00fcni intellekt proqram\u00e7\u0131lar\u0131ndan n\u00fcmun\u0259 g\u0259tirib. Onlar\u0131n m\u00fcr\u0259kk\u0259b tap\u015f\u0131r\u0131qlar\u0131 – m\u0259s\u0259l\u0259n, kodu analiz etm\u0259k v\u0259 funksiyalar\u0131 h\u0259yata ke\u00e7irm\u0259k – bacar\u0131\u011f\u0131na baxmayaraq, bu sisteml\u0259r \u0259ld\u0259 etdikl\u0259ri t\u0259cr\u00fcb\u0259ni saxlam\u0131r v\u0259 bilik toplaya bilmirl\u0259r.<\/p>\n
\u0130nsandan f\u0259rqli olaraq, s\u00fcni intellekt zamanla bacar\u0131qlar\u0131n\u0131 inki\u015faf etdirmir – o, h\u0259r d\u0259f\u0259 \u201c\u00f6yr\u0259nm\u0259y\u0259 s\u0131f\u0131rdan ba\u015flay\u0131r\u201d. \u018fslind\u0259, model \u00fc\u00e7\u00fcn h\u0259r \u201ci\u015f g\u00fcn\u00fc\u201d birinci kimidir. Rafailov h\u0259m\u00e7inin bu c\u00fcr agentl\u0259rin proqramda s\u0259hv ba\u015f verdikd\u0259 i\u015fl\u0259m\u0259y\u0259 davam etm\u0259y\u0259 imkan ver\u0259n try\/except pass kimi konstruksiyalardan istifad\u0259 meylin\u0259 diqq\u0259t \u00e7\u0259kib. O bu hal\u0131n m\u00f6vcud t\u0259lim sisteml\u0259rinin m\u0259hdudlu\u011funu \u0259ks etdirdiyini bildirib: bu modell\u0259r s\u00fcr\u0259tli problem h\u0259llin\u0259 optimalla\u015fd\u0131r\u0131l\u0131b, ancaq anlama v\u0259 \u00f6z\u00fcn\u00fc t\u0259kmill\u0259\u015fdirm\u0259 qabiliyy\u0259tin\u0259 malik deyill\u0259r. Hesablama g\u00fcc\u00fcn\u00fcn daha da art\u0131r\u0131lmas\u0131, h\u0259qiqi \u00fcmumi s\u00fcni intellektin yarad\u0131lmas\u0131na g\u0259tirib \u00e7\u0131xarmayacaq.<\/p>\n
\u201cBiz m\u00f6vcud paradiqmalar \u00e7\u0259r\u00e7iv\u0259sind\u0259 \u00fcmumi s\u00fcni intellekt v\u0259 superintellektl\u0259 bacara bilm\u0259rik\u201d – dey\u0259 Rafailov bildirib. T\u0259dqiqat\u00e7\u0131n\u0131n fikrinc\u0259, miqyas\u0131 art\u0131rmaq \u0259v\u0259zin\u0259 diqq\u0259t keyfiyy\u0259tli m\u0259lumatlara v\u0259 t\u0259lim prosesind\u0259 d\u00fczg\u00fcn m\u00fckafatland\u0131rma strukturuna y\u00f6n\u0259ldilm\u0259lidir. Rafailov \u201cmeta\u00f6yr\u0259nm\u0259\u201d v\u0259 ya \u201c\u00f6yr\u0259nm\u0259yi \u00f6yr\u0259nm\u0259\u201d yana\u015fmas\u0131ndan istifad\u0259 etm\u0259yi t\u0259klif edir. Bu halda modell\u0259r\u0259 bir-birind\u0259n ayr\u0131 tap\u015f\u0131r\u0131qlar deyil, tap\u015f\u0131r\u0131qlarla v\u0259 m\u0259\u015fql\u0259rl\u0259 z\u0259ngin strukturla\u015fd\u0131r\u0131lm\u0131\u015f t\u0259dris kurslar\u0131 t\u0259qdim olunur. Burada m\u00fckafatland\u0131r\u0131lan t\u0259kc\u0259 konkret bir problemi h\u0259ll etm\u0259k deyil, h\u0259m d\u0259 \u00f6yr\u0259nm\u0259 prosesind\u0259ki ir\u0259lil\u0259yi\u015f, abstraksiyalar\u0131 m\u0259nims\u0259m\u0259 v\u0259 uy\u011funla\u015fma qabiliyy\u0259tidir. Bu yana\u015fma t\u0259l\u0259b\u0259nin d\u0259rsliyi m\u0259rh\u0259l\u0259li \u015f\u0259kild\u0259 m\u0259nims\u0259m\u0259sini t\u0259qlid edir v\u0259 art\u0131q AlphaGo kimi sisteml\u0259rd\u0259 \u00f6z s\u0259m\u0259r\u0259liliyini s\u00fcbut edib.<\/p>\n