May ay\u0131n\u0131n ortalar\u0131nda Kaliforniya \u015ftat\u0131n\u0131n Berkli \u015f\u0259h\u0259rind\u0259 30 tan\u0131nm\u0131\u015f riyaziyyat\u00e7\u0131 OpenAI \u015firk\u0259tin\u0259 m\u0259xsus o4-mini adl\u0131 \u00e7at-botun riyazi qabiliyy\u0259tl\u0259rini s\u0131namaq \u00fc\u00e7\u00fcn bir araya g\u0259lib. 2 g\u00fcn \u0259rzind\u0259 onlar bu\u00a0s\u00fcni z\u0259kan\u0131 professor s\u0259viyy\u0259sind\u0259 suallarla s\u0131na\u011fa \u00e7\u0259kibl\u0259r v\u0259 m\u00fc\u0259yy\u0259n edibl\u0259r ki, bu s\u00fcni z\u0259ka sistemi son d\u0259r\u0259c\u0259 \u00e7\u0259tin riyazi m\u0259s\u0259l\u0259l\u0259ri h\u0259ll ed\u0259 bilir v\u0259 art\u0131q d\u00fcnyan\u0131n \u0259n yax\u015f\u0131 magistratura t\u0259l\u0259b\u0259l\u0259rinin \u0259ks\u0259riyy\u0259tini qabaqlay\u0131r. B\u00f6y\u00fck dil modell\u0259ri (BDM) c\u00fcml\u0259d\u0259 n\u00f6vb\u0259ti s\u00f6z\u00fcn n\u0259 olaca\u011f\u0131n\u0131 proqnozla\u015fd\u0131rma qabiliyy\u0259tin\u0259 malikdirl\u0259r. Lakin bir \u00e7ox dig\u0259r BDM-l\u0259rl\u0259 m\u00fcqayis\u0259d\u0259, o4-mini v\u0259 onun Gemini 2.5 Flash\u00a0kimi analoqlar\u0131 daha g\u00fccl\u00fc m\u00f6hk\u0259ml\u0259ndirici \u00f6yr\u0259nm\u0259 il\u0259 x\u00fcsusi m\u0259lumat bazalar\u0131 \u00fcz\u0259rind\u0259 t\u0259liml\u0259ndirilib. Bu yana\u015fma \u00e7at-botu \u0259n\u0259n\u0259vi BDM-l\u0259rl\u0259 m\u00fcqayis\u0259d\u0259 daha m\u00fcr\u0259kk\u0259b riyazi m\u0259s\u0259l\u0259l\u0259r\u0259 d\u0259rind\u0259n n\u00fcfuz ed\u0259 bil\u0259c\u0259k s\u0259viyy\u0259y\u0259 \u00e7atd\u0131rma\u011fa imkan verir.<\/p>\n
OpenAI \u015firk\u0259ti o4-mini modelinin inki\u015faf s\u0259viyy\u0259sini \u00f6l\u00e7m\u0259k \u00fc\u00e7\u00fcn m\u00fcst\u0259qil t\u0259\u015fkilat olan Epoch AI-ya daha \u0259vv\u0259l he\u00e7 vaxt d\u0259rc olunmam\u0131\u015f 300 riyazi m\u0259s\u0259l\u0259d\u0259n ibar\u0259t bir tap\u015f\u0131r\u0131q toplusu haz\u0131rlama\u011f\u0131 tap\u015f\u0131r\u0131b. M\u00fcasir b\u00f6y\u00fck dil modell\u0259ri art\u0131q m\u00fcr\u0259kk\u0259b riyazi suallara do\u011fru cavablar ver\u0259 bilirl\u0259r, lakin Epoch AI t\u0259r\u0259find\u0259n se\u00e7ilmi\u015f bu suallar \u0259n qabaqc\u0131l modell\u0259r \u00fc\u00e7\u00fcn bel\u0259 \u00e7\u0259tin olub – onlar\u0131n he\u00e7 biri 6-dan \u00e7ox tap\u015f\u0131r\u0131q h\u0259ll ed\u0259 bilm\u0259yib, y\u0259ni 2%-lik n\u0259tic\u0259ni a\u015fa bilm\u0259yibl\u0259r. Lakin o4-mini tamamil\u0259 f\u0259rqli, radikal n\u0259tic\u0259 g\u00f6st\u0259rib. Epoch AI t\u0259r\u0259find\u0259n d\u0259v\u0259t olunmu\u015f Elliot Glazer-in r\u0259hb\u0259rlik etdiyi t\u0259dqiqat 2024-c\u00fc ilin sentyabr\u0131nda ba\u015flay\u0131b. o4-mini \u00fc\u00e7\u00fcn n\u0259z\u0259rd\u0259 tutulmu\u015f 300 riyazi tap\u015f\u0131r\u0131q bakalavriat, magistratura v\u0259 t\u0259dqiqat s\u0259viyy\u0259si olmaqla 3 \u00e7\u0259tinlik d\u0259r\u0259c\u0259sin\u0259 b\u00f6l\u00fcn\u00fcb. 2025-ci ilin aprel ay\u0131na q\u0259d\u0259r Glazer m\u00fc\u0259yy\u0259n edib ki, o4-mini bu tap\u015f\u0131r\u0131qlar\u0131n t\u0259xmin\u0259n 20%-ni h\u0259ll ed\u0259 bilir. \u00a0D\u00f6rd\u00fcnc\u00fc s\u0259viyy\u0259 is\u0259 h\u0259tta riyaziyyat professorlar\u0131 \u00fc\u00e7\u00fcn bel\u0259 \u00e7\u0259tin olan tap\u015f\u0131r\u0131qlardan ibar\u0259t olub. T\u0259dqiqatda i\u015ftirak etmi\u015f aliml\u0259r bu suallar\u0131n model t\u0259r\u0259find\u0259n vaxt\u0131ndan \u0259vv\u0259l \u00f6yr\u0259nilm\u0259sinin v\u0259 ya t\u0259sad\u00fcfi s\u0131zman\u0131n qar\u015f\u0131s\u0131n\u0131 almaq m\u0259qs\u0259dil\u0259 m\u0259xfilik sazi\u015fi imzalamal\u0131 olublar. 17-18 may tarixl\u0259rind\u0259 t\u0259cr\u00fcb\u0259nin i\u015ftirak\u00e7\u0131lar\u0131 Berklid\u0259 topla\u015f\u0131b o4-mini-nin son sual blokuna nec\u0259 cavab ver\u0259c\u0259yini yoxlad\u0131lar. 30 alim 6 n\u0259f\u0259rlik qruplara b\u00f6l\u00fcnm\u00fc\u015fd\u00fc. 2 g\u00fcn \u0259rzind\u0259 onlar bir-biril\u0259 yar\u0131\u015faraq el\u0259 tap\u015f\u0131r\u0131qlar haz\u0131rlama\u011fa \u00e7al\u0131\u015fd\u0131lar ki, h\u0259m \u00f6zl\u0259ri onlar\u0131 h\u0259ll ed\u0259 bilsinl\u0259r, h\u0259m d\u0259 \u00e7at-botu \u00e7a\u015fd\u0131rs\u0131nlar. Sonda aliml\u0259r \u00e7at-botu \u00e7\u0131x\u0131lmaz v\u0259ziyy\u0259t\u0259 salan 10 sual tapa bils\u0259l\u0259r d\u0259, s\u00fcni z\u0259kan\u0131n 1 il \u0259rzind\u0259 n\u0259 q\u0259d\u0259r ir\u0259lil\u0259diyi onlar\u0131 heyr\u0259tl\u0259ndirdi. London Riyaziyyat Elml\u0259ri \u0130nstitutunun alimi v\u0259 layih\u0259nin i\u015ftirak\u00e7\u0131lar\u0131ndan biri olan Yang Hui He o4-mini-nin s\u0259viyy\u0259sini \u0259n az\u0131 \u00e7ox savadl\u0131 bir magistrantla m\u00fcqayis\u0259 edib.<\/p>\n
O bildirib ki, \u00e7at-bot h\u0259ll tapma s\u00fcr\u0259ti bax\u0131m\u0131ndan pe\u015f\u0259kar riyaziyyat\u00e7\u0131dan \u0259h\u0259miyy\u0259tli d\u0259r\u0259c\u0259d\u0259 \u00fcst\u00fcnd\u00fcr: orada ki, insan ekspert\u0259 h\u0259ll \u00fc\u00e7\u00fcn h\u0259ft\u0259l\u0259r v\u0259 ya aylar laz\u0131m olard\u0131, model bu i\u015fi c\u0259mi bir ne\u00e7\u0259 d\u0259qiq\u0259y\u0259 yerin\u0259 yetirirdi. S\u00fcni z\u0259kan\u0131n ir\u0259lil\u0259yi\u015fi texnologiyaya heyranl\u0131qla yana\u015f\u0131, narahatl\u0131q da do\u011furur. T\u0259dqiqat i\u015ftirak\u00e7\u0131lar\u0131 o4-mini-nin n\u0259tic\u0259l\u0259rin\u0259 istifad\u0259\u00e7il\u0259rin h\u0259ddind\u0259n art\u0131q g\u00fcv\u0259nm\u0259sind\u0259n narahat olduqlar\u0131n\u0131 ifad\u0259 edibl\u0259r. \u201c\u0130nduksiya il\u0259 s\u00fcbut var, \u0259ksini s\u00fcbut etm\u0259 \u00fcsulu var, bir d\u0259 qorxu il\u0259 s\u00fcbut \u00fcsulu var. \u018fg\u0259r kims\u0259 bir \u015feyi kifay\u0259t q\u0259d\u0259r inamla deyirs\u0259, insanlar sad\u0259c\u0259 etiraz etm\u0259y\u0259 \u00e7\u0259kinirl\u0259r. M\u0259nc\u0259, o4-mini art\u0131q qorxu il\u0259 s\u00fcbutu m\u0259nims\u0259yib – o h\u0259r \u015feyi b\u00f6y\u00fck inamla deyir\u201d – dey\u0259 Yang Hui He bildirib. T\u0259dqiqat\u0131n sonunda i\u015ftirak\u00e7\u0131lar riyaziyyat\u0131n g\u0259l\u0259c\u0259yi bar\u0259d\u0259 d\u00fc\u015f\u00fcnm\u0259y\u0259 ba\u015flay\u0131blar. M\u00fczakir\u0259l\u0259r \u201cbe\u015finci s\u0259viyy\u0259\u201d adland\u0131r\u0131lan m\u0259s\u0259l\u0259l\u0259r\u0259 ke\u00e7ib – bu, h\u0259tta \u0259n yax\u015f\u0131 riyaziyyat\u00e7\u0131lar\u0131n bel\u0259 h\u0259ll ed\u0259 bilm\u0259dikl\u0259ri suallard\u0131r. \u018fg\u0259r s\u00fcni z\u0259ka bu s\u0259viyy\u0259y\u0259 \u00e7atsa, riyaziyyat\u00e7\u0131lar\u0131n rolu k\u0259skin \u015f\u0259kild\u0259 d\u0259yi\u015f\u0259c\u0259k. Bu bax\u0131mdan, riyaziyyat\u0131n g\u0259l\u0259c\u0259k n\u0259sill\u0259r \u00fc\u00e7\u00fcn qorunub saxlanmas\u0131n\u0131n \u0259sas a\u00e7ar\u0131 ali t\u0259hsild\u0259 yarad\u0131c\u0131l\u0131\u011f\u0131n inki\u015faf\u0131 olacaq.<\/p>\n