Ungaqinisekisa kanjani isixazululo sakho se-AI senza lokho okulindele ukuba kwenziwe

IGenai) iyavela ngokushesha – futhi akusekho nje mayelana nezingxoxo ezijabulisayo noma isizukulwane esihlaba umxhwele. 2025 ngunyaka lapho kugxilwe khona ekuguqukeni kwe-AI hype kube yinani langempela. Izinkampani yonke indawo zibheke ngezindlela zokuhlanganisa nokufaka i-genai ngemikhiqizo yazo nezinqubo – ukusebenzela abasebenzisi, ukukhulisa ukusebenza kahle, uhlale uncintisane, futhi ukhule. Futhi sibonga ama-API namamodeli aqeqeshelwe kwangaphambili kusuka kubahlinzeki abakhulu, ukuhlanganisa izizwa lula kunangaphambili. Kepha nakhu ukubanjwa: Ngoba nje ukuhlanganiswa kulula, akusho ukuthi izixazululo ze-AI zizosebenza njengoba kuhlosiwe uma sekuthunyelwe.

Amamodeli wokubikezela akusikho okusha ngempela: njengoba abantu besibikezela izinto iminyaka, ukuqala ngokuhlelekile ngezibalo. Noma kunjalo, IGenai iguqule inkambu yokubikezela ngezizathu eziningi:

• Akunasidingo sokuqeqesha imodeli yakho noma ukuba ngusosayensi wedatha ukwakha izixazululo ze-AI
• I-AI manje kulula ukuyisebenzisa ngokusebenzisa izikhala zokuxoxa kanye nokuhlanganisa nge-API
• Ukuvula izinto eziningi ezazingakwenziwa noma kwakunzima ngempela ukukwenza ngaphambili

Zonke lezi zinto zenza IGenai iyajabulisa kakhulu, kepha futhi iyingozi. Ngokungafani nesoftware yendabuko – noma ngisho nokufunda komshini we-classical – iGenai yethula izinga elisha lokungaphenduki. Awusona ama-logics anqumayo, usebenzisa imodeli eqeqeshiwe kumanani amakhulu wedatha, ngethemba lokuthi izophendula njengoba kudingeka. Ngakho-ke sazi kanjani ukuthi uhlelo lwe-AI lwenza lokho esikuhlose ukukwenza? Sazi kanjani ukuthi kulungele ukuhamba bukhoma? Impendulo ingukuhlola (i-EVALS), umqondo wokuthi sizobe sihlola kulokhu okuthunyelwe:

• Kungani amasistimu we-genai akwazi ukuhlolelwa indlela efanayo ne-software yendabuko noma ngisho nokufunda komshini we-classical (ml)
• Kungani Ukuhlaziywa kubalulekile ukuze uqonde ikhwalithi yohlelo lwakho lwe-AI futhi akayona inketho (ngaphandle kokuthi uthande izimanga)
• Izinhlobo ezahlukene zokuhlola kanye namasu wokuwasebenzisa ekusebenzeni

Noma ngabe ungumphathi womkhiqizo, unjiniyela, noma omunye umuntu osebenzayo noma onesifiso se-AI, ngithemba ukuthi lokhu okuthunyelwe kuzokusiza uqonde ukuthi ucabangani ngokuhlolisisa ngekhwalithi ye-AI (nokuthi kungani oma-e-equals ewukhiye ukufeza leyo mfanelo!).

IGenai ayikwazi ukuhlolwa njenge-software yendabuko- noma i-ml yakudala

Ekuthuthukisweni kwesoftware yendabukoamasistimu alandela imicikilisho enqumayo: Uma kwenzeka, khona-ke kuzokwenzeka – Njalo. Ngaphandle kokuthi kukhona okuqhekeka endaweni yakho noma wethula iphutha kwikhodi … Yisiphi isizathu sokungeza izivivinyo, ukuqapha kanye nezaziso. Ukuhlolwa kweyunithi kusetshenziselwa ukuqinisekisa amabhlogo amancane weKhodi, izivivinyo zokuhlanganisa ukuqinisekisa ukuthi izingxenye zisebenza kahle ndawonye, ​​futhi zibheke ukuthola ukuthi kukhona okuqhuma ukukhiqizwa. Ukuhlola isoftware yendabuko kufana nokubheka ukuthi ngabe umshini wokubala uyasebenza yini. Ufaka 2 + 2, futhi ulindele u-4. Sula futhi enqumayo, kulungile noma akulungile.

Kodwa-ke, i-ML ne-AI yethule ukunganqunyelwe kanye namathuba. Esikhundleni sokuchazwa kokuziphatha ngokusobala ngemithetho, siqeqesha amamodeli ukuze afunde amaphethini avela kwimininingwane. E-AI, uma kwenzeka, okuphumayo akusekho okuhlangene y, kepha ukubikezela ngezinga elithile okungenzeka, ngokusekelwe kulokho okufundwe yimodeli ngesikhathi sokuqeqeshwa. Lokhu kungaba namandla amakhulu, kepha futhi kwazisa ukungaqiniseki: okokufaka okufanayo kungenzeka kube nemiphumela ehlukile ngokuhamba kwesikhathi, imiphumela ebonakalayo empeleni ingahle ingalungile, indlela yokuziphatha engalindelekile yezimo ezingandile ingahle ivele …

Lokhu kwenza ukuthi izivivinyo zendabuko zisondela azanele, azize zibe khona ngezikhathi ezithile. Isibonelo sokubala sisondela ekuzameni ukuhlola ukusebenza komfundi ekuhlolweni okuvulelekile. Ngombuzo ngamunye, nezindlela eziningi ezingenzeka zokuphendula umbuzo, kuyimpendulo enikezwe okulungile? Ingabe ngaphezu kwezinga lolwazi umfundi okufanele abe nalo? Ngabe umfundi wakwenza konke kube yikhona kodwa kuzwakale kukholisa kakhulu? Njengezimpendulo ezivivinyweni, Izinhlelo ze-AI zingahlaziywa, kepha zidinga indlela ejwayelekile futhi eguquguqukayo yokuzivumelanisa nokufakwa okuhlukile, izimo kanye nokusebenzisa amacala (noma izinhlobo zezivivinyo).

Emshini wendabuko wokufunda (ML), ukuhlolwa sekuvele kuyingxenye esungulwe kahle ye-Project Lifecycle. Ukuqeqesha imodeli kumsebenzi omncane njengokuvunyelwa kwemalimboleko noma ukutholwa kwezifo kuhlanganisa isinyathelo sokuhlola – ukusebenzisa amamethrikhi njengokunemba, ukucacisa ukuthi imodeli isebenza kahle yini ukuthumela phambili ekuthuniselweni. Ku-genai lokhu kuvame ukuguquka: amaqembu asebenzisa amamodeli asevele aqeqeshiwe futhi asevele edlulile ukuhlolwa kwenhloso ejwayelekile kuwo wonke ngaphakathi kuMhlinzeki Wemodeli. Lezi zinhlobo zilungile kakhulu emisebenzini ejwayelekile – njengokuphendula imibuzo noma ukubhala ama-imeyili – kunengozi yokubahlula ngecala lethu lokusebenzisa elithile. Kodwa-ke, kubalulekile ukucela “Ingabe le modeli emangalisayo ilungile ngokwanele ngecala lami lokusebenzisa?“. Kulapho ukuhlolwa kungena khona – Ukuhlola ukuthi iziqubulo noma izizukulwane zilungele icala lakho elithile lokusebenzisa, umongo, okokufaka kanye nabasebenzisi.

Kunomunye umehluko omkhulu phakathi kwe-ML neGenai: Izinhlobonhlobo nobunzima bemiphumela yemodeli. Asisabuyi abuyisele amakilasi namathuba (njengokuthi kungenzeka iklayenti lizobuyisa imali mboleko), noma izinombolo (njengentengo yendlu ebikezelwe ngokusekelwe kwizimpawu zayo). Izinhlelo zeGenai zingabuyisa izinhlobo eziningi zokuphuma, ubude obuhlukile, ithoni, okuqukethwe, nefomethi. Ngokufanayo, la mamodeli awadingi okokufaka okuhleliwe futhi anqunywe kakhulu, kepha ajwayele ukuthatha cishe noma yiluphi uhlobo lokufaka – umbhalo, izithombe, nomsindo noma ividiyo. Ngakho-ke kuhlaziya kube nzima kakhulu.

Kungani U-Evils Akakhethi (Ngaphandle kokuthi Uthande Izimanga)

U-Evals kukusiza ukukala ukuthi ngabe uhlelo lwakho lwe-AI empeleni lusebenza ngendlela you swela Kuzo, ukuthi uhlelo lukulungele ukuhamba bukhoma, futhi uma nje luphilisa lulokhu luqhubeka nokwenza njengoba bekulindelekile. Ukudiliza KUNGANI KUNGANI KUBALULEKILE:

• Ukuhlolwa kwekhwalithi: U-Evals uhlinzeka ngendlela ehlelekile yokuqonda ikhwalithi yokuqagela kwakho kwe-AI noma imiphumela yokuphuma nokuthi izohlanganisa kanjani ohlelweni oluphelele kanye nasesimweni sokusetshenziswa. Ingabe izimpendulo zinembile? Kuwusizo? Uhambisane naye? Efanele?
• Iphutha Iphutha: Ukuhlola kusiza ukunciphisa amaphesenti, izinhlobo, kanye nobukhulu bemaphutha. Kukangaki izinto zingahambi kahle? Hlobo luni lwamaphutha avela kaningi (isib. Positive wamanga, ama-hallucinations, ukufometha amaphutha)?
• Ukuncipha kobungozi: Ikusiza ukubona futhi uvikele isimilo esilimazayo noma esinengozi ngaphambi kokuthi ifinyelele kubasebenzisi – Ukuvikela inkampani yakho kusuka engcupheni yezobungozi, izindaba zokuziphatha, kanye nezinkinga ezilawulwa.

I-AIRETION AI, ngobudlelwano bayo bokufaka kwamahhala bokufaka kanye nesizukulwane eside sombhalo, yenza ukuhlolwa okubaluleke kakhulu futhi kuyinkimbinkimbi. Lapho izinto zingahambi kahle, zingahle zingahambi kahle. Sonke sibone ama-headlines mayelana nezingxoxo ezinikeza izeluleko eziyingozi, amamodeli akhiqiza okuqukethwe okubandlululayo, noma amathuluzi we-AI ahlehlisa amaqiniso wamanga.

“I-AI ayisoze yaphelela, kepha ngezinamba unganciphisa ubungozi bokuthi amahloni – okungabiza imali, ukuthembeka, noma umzuzu wegciwane ku-Twitter.“

Uyichaza kanjani isu lokuhlola?

Ngakho-ke sikuchaza kanjani ukuhlolwa kwethu? U-Evals awekho ngosayizi owodwa – konke. Ziyancike – zincike kakhulu futhi kufanele zivumelane nezinhloso ezithile zohlelo lwakho lokusebenza lwe-AI. Uma wakhe injini yokusesha, ungakhathalela ukuhambisana nomphumela. Uma kuyingxoxo, ungahle ukhathalele usizo nokuphepha. Uma kuyisisekelo, kungenzeka ukuthi unendaba nokunemba nokunemba. Ngezinhlelo ezinezinyathelo eziningi (njengohlelo lwe-AI ezenza usesho, zibeka phambili imiphumela bese zikhiqiza impendulo) kuvame ukudingekile ukuhlola isinyathelo ngasinye. Umqondo lapha ukukala uma isinyathelo ngasinye sisiza ukufinyelela i-General Accicic Metric (futhi ngokusebenzisa lokhu kuqonda ukuthi ungagxila kuphi nokuthuthuka).

Izindawo zokuhlola ezijwayelekile zifaka:

• Ukunemba nokuhlaziywa: Ingabe imiphumela inembe ngokuphelele? Ngabe bakwenza izinto?
• Ukuhambisana: Ingabe okuqukethwe kuqondaniswe nombuzo womsebenzisi noma umongo onikezwe?
• ukuphepha, ukukhetha, kanye nobuthi
• Ifomethi: Ingabe imiphumela ngefomethi elindelekile (isib., Json, Call Offer Work Call)?
• Ukuphepha, ukukhetha nobuthi: Ingabe uhlelo lukhiqiza okulimazayo, okunengqondo, noma okunobuthi?

Amamethrikhi athile athile. Isibonelo emisebenzini yokuhlukanisa izindlela ezinjengokunemba nokunemba, ukufingqa imisebenzi i-Rouge noma i-Bleu, futhi ku-Code Generals Task Tasks regex kanye nokubulawa kwephutha.

Ngabe uhlanganisa kanjani ubumbano?

Lapho usuyazi ukuthi ufuna ukukala ini, isinyathelo esilandelayo siklama amacala akho wokuhlola. Lokhu kuzoba iqoqo lezibonelo (izibonelo eziningi ezisezingeni elingcono, kepha zihlala zilinganisa nezindleko) lapho unayo:

• Isibonelo sokufaka: Ukufakwa okungokoqobo kohlelo lwakho kanye ekukhiqizeni.
• Ukukhishwa okulindelekile (Uma kusebenza): Iqiniso lomhlaba noma isibonelo semiphumela efiselekayo.
• Indlela yokuhlola: Indlela yokushaya amagoli ukuhlola umphumela.
• Amamaki noma adlule / wehluleka: Imetric ehlanganisiwe ehlola icala lakho lokuhlola

Ngokuya ngezidingo zakho, isikhathi, kanye nesabelomali, kunamasu ambalwa ongawasebenzisa njengezindlela zokuhlola:

• Ama-Stistical Scorers afana I-Bleu, Rouge, i-Meteor, noma ukufana kwe-cosining phakathi kokushumeka – okuhle ngokuqhathanisa umbhalo okhiqizwayo ku-Reference Outents.
• Amamethrikhi wendabuko ml afana Ukunemba, ukunemba, ukukhumbula, kanye ne-AUC – okungcono kakhulu ekuhlukaniselweni ngemininingwane ebelwe.
• I-LLM-A-A-JURADI Sebenzisa imodeli enkulu yolimi ukukala ukuphuma (isib.Ingabe le mpendulo iyiqiniso futhi iyasiza?“). Isebenziseka kakhulu lapho idatha ebizwayo ayitholakali noma lapho kuhlaziywa isizukulwane esivulelekile.

Ukudabuka Okusekelwa Ikhodi Sebenzisa i-regex, imithetho enengqondo, noma ukwenziwa kwecala lokuhlola ukuze uqinisekise amafomethi.

Uyisonge

Ake sibuyisele konke ngesibonelo sokhonkolo. Cabanga ukuthi wakha uhlelo lokuhlaziya imizwa ukusiza iqembu lakho lokuxhaswa kwamakhasimende kuqala ama-imeyili angenayo.

Umgomo ukuqiniseka ukuthi imilayezo ephuthumayo noma engemihle kakhulu ithola izimpendulo ezisheshayo – ukunciphisa ukunciphisa ukukhungatheka, ukuthuthukisa ukwaneliseka, nokuncipha kwe-churn. Leli yicala elisebenziseka kalula, kepha ngisho nasesihlelweni esinjengalesi, ngokuphuma okulinganiselwe, izindaba ezisezingeni elifanele: Izibikezelo ezimbi zingaholela kuma-imeyili abeka phambili ngezikhathi ezithile, okusho ukuthi iqembu lakho lichitha isikhathi.

Ngakho-ke wazi kanjani ukuthi isixazululo sakho sisebenza nekhwalithi edingekayo? Uhlola. Nazi ezinye izibonelo zezinto ezingase zibe zifanelekile ukuhlola kuleli cala elithile lokusebenzisa:

• Ukuqinisekiswa kwefomethi: Ingabe imiphumela yocingo lwe-LLM ukubikezela imizwa ye-imeyili ebuyiselwe kufono ye-json elindelekile? Lokhu kungahlaziywa ngamasheke asuselwa kukhodi: i-regex, ukuqinisekiswa kwe-schema, njll.
• Ukunemba kokucaciswa kokuqonda: Ingabe uhlelo luhlukanisa kahle imichilo emibhalweni emifushane – emfushane, isikhathi eside, esilimini esiningi? Lokhu kungahlaziywa ngemininingwane enelebuli kusetshenziswa amamethrikhi wendabuko we-ML – noma, uma amalebula engatholakali, esebenzisa i-LLM-AS-A-A-JURADI.

Lapho ikhambi libukhoma, ungafuna ukufaka amamethrikhi ahlobene kakhulu nomthelela wokugcina wesisombululo sakho:

• Ukuphumelela kokuphumelela: Ingabe ama-ejenti asekelayo empeleni aqondiswa kuma-imeyili abucayi kakhulu? Ingabe ukubeka phambili kuhambisana nomthelela webhizinisi owufunayo?
• Umthelela Wokugcina Webhizinisi Ngokuhamba kwesikhathi, ingabe lolu hlelo lunciphisa izikhathi zokuphendula izikhathi, ukwehlisa i-Churn yamakhasimende, kanye nokwenza ngcono izikolo zokweneliseka?

E-EVALLS iyisihluthulelo sokuqinisekisa ukuthi sakha izinhlelo eziwusizo, eziphephile, ezibalulekile, ezilungele ukusetshenziswa kwe-AI ekukhiqizeni. Ngakho-ke, noma ngabe usebenza nge-chalbot elula noma i-chatbot evulekile evulekile, thatha isikhathi ukuchaza ukuthi yikuphi “okuhle ngokwanele” kusho (ubuhle obusezingeni eliphezulu) – bese wakhe amakhwalithi asebenzayo azungeze ukukala!

Ukunqubekela phambili

[1] Umkhiqizo wakho we-AI udinga ama-EVALL, HAMEL HUSAIN

[2] Amamethrikhi we-LLM Ukuhlola: Umhlahlandlela wokugcina wokuhlola we-LLM, othembela ai

[3] Ukuhlola ama-agents we-AI, ngokudambisa.ai + adane