Zeziphi iimbono eziziswa yimpumelelo yeDeepSeek kushishino lwezothutho lukaloliwe?
AS Inkampani yaseTshayina egxile ekufezekiseni i-AGI, iDeepSeek iphumelele kakhulu kwezobuchwepheshe kunye noyilo lwemodeli yeshishini kwicandelo lobukrelekrele bokwenziwa kunye neemodeli ezinkulu zibonelela ngengqiqo ebanzi kushishino lwezothutho lukaloliwe. Olu hlalutyo lulandelayo luza kwenziwa ngokwembono yetekhnoloji, ulawulo, kunye nesicwangciso:
1, iTekhnoloji yoQhagamshelo: Utshintsho olukrelekrele oluqhutywa yi-AI loThutho lweZitimela
Uguquko olutshintshayo lweshedyuli: Ukuphuhlisa imodeli yeshedyuli yeendlela ezininzi esekelwe kuyilo lweTransformer ukuze kufezekiswe ukwenziwa ngcono kweshedyuli yesitimela kwinqanaba le-millisecond. Emva kokusebenzisa ubuchwepheshe obufanayo kwiTokyo Metro, umthamo weyure ende unyuke nge-23% kwaye ukusetyenziswa kwamandla kwehle nge-12%. Ityala: Iprojekthi yeCrossrail eLondon idibanisa inkqubo yedijithali, elungisa ngokuzenzekelayo isicwangciso sokudibanisa ngokusebenzisa ukuqikelela ukuhamba komkhweli ngexesha langempela, kwaye iphucula ukusebenza kakuhle kokusingatha ukulibaziseka ngequbuliso nge-40% ngo-2023. Uphuhliso ekugcinweni kwangaphambili: Ukuphuhlisa igrafu yolwazi lwesimo sethrekhi, ukudibanisa idatha yesenzi sokufuduka kwelaser kunye neerekhodi zokugcinwa kwembali. Emva kovavanyo lweShenzhen Metro, ukuchaneka kokubikezela ukuguqulwa kwejometri yethrekhi kufikelele kwi-98.7%, kwaye iindleko zokugcinwa zehle nge-35%. I-Deutsche Bahn DB isebenzisa itekhnoloji yokuqaphela i-voiceprint ukubona ukungahambi kakuhle kwereyile yamavili ngokusebenzisa ii-microphone arrays ezikwibhodi, ngesantya sesilumkiso se-89% kwiintsuku ezili-14 kwangaphambili.
2, Ukwakhiwa ngokutsha kwendlela yokusebenza: ukukhutshwa kwexabiso leeasethi zedatha
Ukwembiwa kwexabiso lokuhamba komkhweli: Yakha imodeli yenethiwekhi ye-neural yegrafu yendawo ukuze iguqule iindlela zokuhamba komkhweli zibe ziimephu zobushushu bokuhamba korhwebo. Ngokusekelwe koku, iShanghai Hongqiao Hub ilungise uyilo lweevenkile, nto leyo ebangele ukwanda kwe-19% kwingeniso engeyiyo itikiti. Imodeli ethi "Railway+Property" yeHong Kong MTR inyuse izinga lempumelelo yophuhliso lweprojekthi ye-TOD ngamanqaku angama-27 eepesenti ngohlalutyo lwedatha yokuhamba. Ulawulo lwamandla olukrelekrele: Iphuhlise ii-algorithms zolawulo lokufunda oluqinisiweyo kwiinkqubo zonikezelo lwamandla okutsala, kwaye yonyusa izinga lokusetyenziswa kwamandla okuqhobosheka okuvuselelayo kwiBeijing Metro Line 10 ukusuka kwi-65% ukuya kwi-82%. Ukwenziwa ngcono kwenkqubo yokudonsa amandla yokugcina i-photovoltaic energy storage Tokyo Metro, kufezekisa izinga lokusetyenziswa kwe-photovoltaic eliphakathi kwemihla ngemihla le-91.2% ngo-2024.
3, Utshintsho kwiZiko: Ukwakha iNkqubo yeNdalo eVuselelekayo
Uhlaziyo lweNdalo yoPhando noPhuhliso: Seka "iQonga eliVulekileyo leeModeli ezinkulu zeSitimela" ukutsala ababoneleli bezixhobo abangaphezu kwama-300 ukuba bajoyine, ukunciphisa ixesha eliphakathi lokuphendula ukuxilongwa kwempazamo ukusuka kwimizuzu engama-45 ukuya kwimizuzu esi-8. I-Guangzhou Metro kunye ne-SenseTime Technology baseke ilebhu edibeneyo, eyonyuse ukusebenza kakuhle kokuhlolwa kwenethiwekhi yoqhagamshelwano ngamaxesha ali-15 kwaye yanciphisa izinga le-alamu yobuxoki ukuya kwi-0.3%. Ukuguqulwa kwesakhiwo setalente: Ukuphumeza inkqubo yoqeqesho lwetalente edibeneyo ye-"AI+Rail", inani leenjineli zedatha eChengdu Metro linyuke ukusuka kwi-3% ukuya kwi-12%, kwaye imveliso yelungelo lomenzi wechiza leqela le-algorithm inyuke kahlanu. I-SMRT eSingapore imisele isikhundla seGosa eliyiNtloko le-AI ukulungelelanisa nokukhuthaza iiprojekthi ezingama-23 zotshintsho olukrelekrele.
4, Uphuculo lweqhinga: ukuchaza ngokutsha ixabiso lokuhamba ngololiwe
Ukuzinzisa iMobility as a Service (MaaS): Ukuphuhlisa injini yesigqibo sothutho oluneendlela ezininzi edibanisa idatha evela kwiindlela ezili-17 zothutho. Iqonga leHangzhou "Zhe Li Chang Xing" linciphise ixesha eliqhelekileyo lokudlulisa imowudi enqamlezileyo ngemizuzu engama-22. Inkqubo yetyala lokuhamba elilingwayo kwiXiong'an New Area isebenzise imodeli "yokukhwela kuqala, uhlawule kamva" ngokusekelwe kwidatha yokuziphatha kwabakhweli, kunye nezinga lokuqokelelwa kwamatikiti elonyuswe laya kwi-99.8%. Ukwakhiwa kwenkqubo yedijithali yewele: Misela inkqubo epheleleyo yolawulo lweeasethi ze-3D ukuphucula ukuchaneka kolawulo lomjikelo wobomi bezixhobo ukuya kwinqanaba le-millimeter. Inkqubo yokusebenza ekrelekrele kunye nokugcinwa kololiwe ohamba ngesantya esiphezulu waseBeijing Zhangjiakou inciphisa umthwalo womsebenzi wokuhlolwa ngesandla nge-73%. I-digital twin yaseDubai Metro ifezekisa ukwenziwa kwe-virtualization kwe-emergency drills, inyusa isantya sokwenza izicwangciso zokuphendula emergency ngamaxesha angama-40.
5, Ukuthintela nokulawula umngcipheko: isiqinisekiso esithembekileyo kwixesha lobukrelekrele
Ukuphuculwa koKhuseleko loKhuseleko: Kwenziwe inethiwekhi yokuvelisa echaseneyo yokufumanisa ukungena, yathintela ngempumelelo i-99.97% yohlaselo lwenkqubo yolawulo lwemizi-mveliso ngesantya se-alamu yobuxoki esilawulwa ngaphantsi kwe-0.02%. Sisebenzisa itekhnoloji yokufunda edityanisiweyo ukufezekisa ukwabelana ngedatha yokhuseleko oluphakathi kwezixeko, ixesha lokuhlaziya ubukrelekrele bezoyikiso lincitshisiwe ukusuka kwiiyure ezingama-72 ukuya kwimizuzu eli-15. Isakhelo solawulo lokuziphatha: Seka inkqubo yovavanyo lokutolikwa kwesigqibo se-AI, enamanqaku okucaca angama-4.8/5 kwii-algorithms eziphambili zenkqubo. Phuhlisa isicwangciso sokukhusela ulawulo lwedatha ukuze kufezekiswe imigangatho yesatifikethi se-GDPR yokucubungula idatha yobumfihlo yabakhweli.
Imbono yexesha elizayo: Ishishini lezothutho lukaloliwe lijongene notshintsho oluvela kwingcamango ethi "ukwenziwa koomatshini → ukwenziwa kwedijithali → ubukrelekrele". Indlela yokusebenza ye-DeepSeek ibonise ukuba impumelelo yezobuchwepheshe kufuneka ikhuthazwe ngaxeshanye notshintsho lwentlangano kunye nohlengahlengiso lwendalo. Kucetyiswa ukuba eli shishini liseke iziko lokugqwesa le-AI, ligxile kwimpumelelo kwiinkalo eziphambili ezifana nolawulo oludibeneyo lwee-arhente ezininzi kunye nokwenza ngcono i-quantum computing. Kwangaxeshanye, inkqubo yolawulo lwe-AI kufuneka iphuculwe ukuze kufezekiswe utshintsho kwinqanaba leshishini phantsi kwesiseko sokhuseleko kunye nolawulo. Ngokwezibikezelo ezivela kwi-Korea Railway Research Institute (KRRI), ubukrelekrele obupheleleyo bunokunciphisa iindleko zokusebenza zothutho lukaloliwe ngama-38% kwaye bonyuse amandla enkonzo ngama-55%, okunokuba yindlela yokuguquka kwesizukulwana esilandelayo sothutho lukaloliwe olukrelekrele.