Izinkampani ezibucayi zezinkampani ezibucayi zenza lapho zihlanganisa i-AI / ml ezinqubweni zazo

esichithe umsebenzi wami osebenza ngaphesheya kwezimboni ezahlukahlukene, kusukela ekubhaleni okuncane ezinkampanini zomhlaba wonke, ezinkampanini ze-AI-Fight Tech kuya emabhange alawulwa kakhulu. Kuyo yonke le minyaka, ngibone imizamo eminingi ye-AI ne-ML iphumelela, kodwa futhi ngibone inombolo exakayo yehluleka. Izizathu zokwehluleka zivame ukukwenza kancane ukukwenza ngama-algorithms. Imbangela yezimpande cishe njalo izinhlangano zisondela e-AI.

Lokhu akulona uhlu lokuhlola, kanjani-kuzandla, noma uhlu lwemithetho enzima futhi esheshayo. Ukubuyekezwa kwamaphutha ajwayelekile engikutholile, kanye nokuqagela okuthile mayelana nokuthi kungani benzeka, nokuthi ngicabanga ukuthi bangagwenywa kanjani.

1. Ukuntuleka kwesisekelo sedatha eqinile

Uma kungekho mmpofu, noma idatha encane, kaningi ngenxa yokuvuthwa okuphansi kwezobuchwepheshe, amaphrojekthi we-AI / ML ahlelwe ukwehluleka. Lokhu kwenzeka kaningi lapho izinhlangano zikha amathimba we-DS / ML ngaphambi kokuba asungule imikhuba yobunjiniyela eqinile yedatha.

Ngangithi umphathi uthi kimi, “ama-spreadsheets awenzi imali.” Ezinkampanini eziningi, nokho, okuphambene ngqo: “Ama-Spreadsheets” yilona kuphela ithuluzi elingasunduza inzuzo phezulu. Ukwehluleka ukwenza kanjalo kusho ukuba izisulu ze-aphorism zakudala: “udoti phakathi, udoti ngaphandle.”

Bengihlala ngisebenza enkampanini yokulethwa kokudla kwesifunda. Amaphupho wethimba le-DS abekade eliphakeme kakhulu: Izinhlelo zokufunda ezijulile, gen AI, njll. Kepha idatha yakudala yezakhiwo ezindala ngakho-ke amaseshini anazo ngokuthembekile ngoba bekungekho i-ID eyodwa yokhiye; Ama-ID e-Restaurant Dish ajikeleze njalo emavikini amabili, ngakho-ke akunakwenzeka ukuthi acabange ngokuphepha lokho amakhasimende a-oda ngempela. Lokhu nezinye izingqinamba eziningi kusho ukuthi yonke iphrojekthi yayingama-70% Worparound. Akunasikhathi noma izinsizakusebenza zezixazululo ezinhle kakhulu. Kepha kube nedlanzana lazo, akekho noyedwa wawo wamaphrojekthi owayeveze eminye imiphumela kungakapheli unyaka ngoba wakhulelwa kususelwa kudatha ebingathenjwa.

Ukudla okudlela endlini: Tshala imali ebunjiniyela bedatha kanye nokuqapha kwekhwalithi yedatha ngaphambi kwe-ML. Gcina kuqondile. Ukuwina kwaqala kanye nezithelo ezilengiswe phansi “akudingi idatha esezingeni eliphakeme, kepha i-AI ngokuqinisekile izokwenza.

2. Alikho icala elicacile lebhizinisi

I-ML ijwayele ukwenziwa ngoba iyathandeka kunokuba ixazulule inkinga yangempela, ikakhulukazi inikezwe i-LLM ne-Agentic AI Hype. Izinkampani zakha ukusetshenziswa kwamacala azungeze ubuchwepheshe kunokwenye indlela ezungeze indlela, eqeda ukwakha izixazululo eziyinkimbinkimbi noma ezingafuneki.

Cabanga ngomsizi we-AI ohlelweni lokukhokha olukhokhelwayo lapho amakhasimende acindezela kuphela izinkinobho ezintathu, noma umhumushi we-AI wamadeshibhodi lapho isixazululo kufanele senze amadeshibhodi aqondakale. Ukucinga okusheshayo kwe-Google kwezibonelo zangasizi we-AI Ohlulekile kuzothuthukisa izimo eziningi ezinjalo.

Isimo esisodwa esinjalo empilweni yami yokusebenza sasiyiphrojekthi yokwakha umsizi otholakele ekutholeni indawo yokudlela nokubhuka (umhlanganisi wokudlela, ake sithi). I-LLMS yayingukufutheka konke, futhi kwakukhona i-fomo kusuka phezulu. Banqume ukuthuthukisa insizakalo ephephile engcono kakhulu nomsizi wengxoxo obhekene nomsebenzisi. Umsizi angaphakamisa izindawo zokudlela ngokuya ngezicelo ezinjengokuthi “ngibonise izindawo ezinhle ngezaphulelo,” “Ngifuna isidlo sakusihlwa sentombi yami,” noma “thola izindawo ezinobungane.”

Kwathi ngonyaka wachithwa ukuthuthuka yithimba: Amakhulu ama-Scenarios aklanywe, ama-Guardrails ahlelwe, i-Backlend yabeka i-bulletoof. Kepha umongo wendaba ukuthi lo msizi akazange axazulule noma yimaphi amaphuzu obuhlungu bangempela bomsebenzisi. Amaphesenti amancane abasebenzisi aze azama ukuwasebenzisa futhi phakathi kwabo kuphela inani lezibalo lezinto ezingelutho eliholele ekubhukeni. Iphrojekthi yashiywa kusenesikhathi futhi ayizange ilinganiselwe kwezinye izinsizakalo. Uma iqembu lase liqale ngokuqinisekiswa kwecala lokusebenzisa esikhundleni sezici zomsizi, isiphetho esinjalo sasingatholakali.

Ukudla okudlela endlini: Qala ngenkinga njalo. Qonda iphuzu lobuhlungu ngokujulile, nika inani layo ngamanani, bese uqala uhambo lokuthuthuka.

3. Ukujaha ubunzima ngaphambi kokungena izisekelo

Imiphakathi eminingi igxuma enguqulweni yakamuva ngaphandle kokuma ukubona ukuthi ngabe izindlela ezilula zizokwanela. Usayizi owodwa awuhambelani konke. Indlela ekhuphukayo, ukuqala okulula nokukhuphula njengoba kudingeka, cishe njalo kuphumela kuma-ROGER ROI. Kungani kukwenza kube nzima ukwedlula ukuthi kudingeka kube nini lapho kutholwa khona i-regression, amamodeli aqeqeshwe ngaphambi kwesikhathi, noma ama-hearistis azolahla? Ukuqala okulula kuhlinzeka ngokuqonda: ufunda ngenkinga, thola ukuthi kungani ungaphumeleli, futhi ube nesisekelo esizwakalayo sokuya ngokuhamba kwesikhathi.

Ngisebenzise iphrojekthi yokwakha iphrojekthi yokuqamba iwijethi ye-Shortcut ekhasini lasekhaya le-Multi-Service App efaka ukugibela okuhilelekile. Umqondo wawulula: Qagela ukuthi umsebenzisi wethule uhlelo lokusebenza ukucela ukugibela, futhi uma kunjalo, abikezela ukuthi kungenzeka yini ukuthi umsebenzisi angakubhukha ukuze umsebenzisi akwazi ukuwubhukha ngokuthinta okukodwa. Ukuphathwa kwanquma ukuthi isixazululo kumele sibe yinethiwekhi ye-neural futhi kungenzeka kube yilutho olunye. Ezinyangeni ezine zokuvela kobuhlungu ngemuva kwalokho, sathola ukuthi ukubikezela kwenziwa kahle ngokumangazayo mhlawumbe mhlawumbe ama-10% abagibeli abanomlando ojulile wokugcina ukugibela. Ngisho nakwabo, izibikezelo zazimbi. Futhi inkinga ekugcineni yahlelwa ngobusuku obubodwa ngeqoqo lemithetho yebhizinisi. Izinyanga zemizamo yokuchitha kungenzeka ukuthi zivinjelwe uma inkampani iqale ngokulondolozwa.

Ukudla okudlela endlini: Uhambe ngaphambi kokugijima. Sebenzisa ubunzima njenge-resort yokugcina, hhayi indawo yokuqala.

4. Nqamula phakathi kwamaqembu e-ML nebhizinisi

Ezinhlanganweni eziningi, isayensi yedatha yisiqhingi. Amaqembu akha izixazululo ezimangalisa ubuchwepheshe ezingalokothi zithole ukukhanya kosuku ngoba azixazululi izinkinga ezifanele, noma ngoba ababambiqhaza bebhizinisi ababethembi. Ukubuyela emuva akungcono I-Equilibrium yimpendulo. I-ML iyachuma kahle lapho kungumsebenzi ngokubambisana phakathi kwabachwepheshe besizinda, onjiniyela, kanye nabenzi bezinqumo.

Ngikubonile lokhu kaningi ezinkampanini ezinkulu ezingezona-zendabuko. Babona ukuthi i-AI / ML inamandla amakhulu futhi isethelwe “Ai Labs” noma izikhungo zobuhle. Inkinga le ama-labs ngokuvamile asebenza ngokuhlukaniswa okuphelele ebhizinisini, futhi izixazululo zawo azivamile ukutholwa. Ngisebenzele ibhange elikhulu elalinelebhu nje elinjalo. Kwakunabangoti abanolwazi kakhulu lapho, kepha abakaze bahlangane nababambiqhaza bebhizinisi. Okubi kakhulu, ilabhorethri yasungulwa njengezifiso zokuma zodwa, kanye nemininingwane yokushintshana kwakungenzeki. I-Firm ibingeyona eyanele emsebenzini welebhu, owagcina ukuya emaphepheni okucwaninga ngezifundo kodwa hhayi ezinqubweni zangempela zenkampani.

Ukudla okudlela endlini: Gcina imizamo ye-ML ihambisane kahle nezidingo zebhizinisi. Sebenzisana kusenesikhathi, xhumana kaningi, bese ushintsha nababambiqhaza, noma ngabe kunciphisa intuthuko.

5. Ukungazinaki ama-MLops

Imisebenzi ye-Cron nemibhalo eyisisekelo izosebenza ngezinga elincane. Lokho kusho, njengezikali eziqinile, le yindlela yokupheka yenhlekelele. Ngaphandle kwama-mlops, ama-tweaks amancane adinga ukubandakanyeka konjiniyela bokuqala zonke izinyathelo zendlela, futhi amasistimu abhalwe kabusha ngokugcwele kaninginingi.

Ukutshala imali kusenesikhathi kuma-MLOPS ukhokha ngokwezifiso. Akukona nje ubuchwepheshe, kepha ukuba nesiko elizinzile, elinamandla futhi elizinzile le-ML.

Ukutshala imali kuma-MLops kusenesikhathi kukhokha ngokungenangqondo. Akukona nje ngobuchwepheshe; Kumayelana nokwakha isiko lokuthembekile, elinezinkozo, futhi elisebenzayo ml. Ungavumeli isiyaluyalu sehlele. Sungula izinqubo ezinhle, amapulatifomu, nokuqeqeshwa ngaphambi kwamaphrojekthi we-ML agijima endle.

Ngisebenze nge-telecom subsididies firm ye-adtech. Ipulatifomu ibisebenzela ukukhangisa nge-Intanethi futhi kwakuyimali enkulu kunazo zonke yenkampani. Ngoba bekuyintsha (kuphela ubudala) isixazululo se-ML sasihlasela ngokulangazelela. Amamodeli avele afakwe ku-C ++ kuphela futhi afakwe kwikhodi yomkhiqizo ngunjiniyela owodwa. Ukuhlanganiswa kwenziwa kuphela uma lowo njini bekhona, amamodeli awakaze alandelwe, futhi uma umbhali wokuqala ashiywe, akekho noyedwa owayenenkomba yokuthi basebenza kanjani. Uma unjiniyela we-Shift waphinde wahamba, yonke ipulatifomu ibingahle ishiye unomphela. Ukuvezwa okunjalo kungenzeka kuvinjelwe ama-mlops amahle.

6. Ukuntuleka kokuhlolwa kwe-A / B

Amanye amabhizinisi agwema ukuhlolwa kwe-A / B ngenxa yobunzima nokukhetha izipele ezingemuva noma ekuhlolweni esikhundleni salokho. Evumela amamodeli amabi ukufinyelela ukukhiqizwa. Ngaphandle kwesikhulumi sokuhlola, umuntu akakwazi ukuthi yimaphi amamodeli asebenza. Uhlaka olufanele lokuhlola luyadingeka ukuthuthuka okukhona, ikakhulukazi esikalini.

Okuthambekele ekubambeni ukwamukelwa ngumuzwa wobunzima. Kepha inqubo yokuhlola eqondile, ehambisane kahle, ingasebenza kahle ezinsukwini zokuqala futhi ayidingi ukutshalwa kwezimali okukhulu. Ukuqondanisa nokuqeqeshwa kuyizikhiye ezinkulu kakhulu.

Endabeni yami, ngaphandle kwendlela ezwakalayo yokulinganisa umthelela womsebenzisi, kungukuthi imenenja ingayithengisa kahle kangakanani. Ama-Pitches amahle axhaswe, avikelwe ngobuqotho, futhi kwesinye isikhathi agcina ngisho noma izinombolo zinciphisa. Amamethrikhi aphethwe ngokuqhathanisa izinombolo zokuqalisa zangaphambi / zokuthumela. Uma bekhulisa, iphrojekthi impumelelo, yize kwenzeka kanjalo. Kumafemu akhulayo, kunezigidi zamaphrojekthi we-Subber afihlwe ngemuva kokukhula okuphelele ngoba akukho ukuhlolwa kwe-A / B ukuze ahlukanise amathuba okuqhubeka nokwehluleka.

Ukudla okudlela endlini: Dala umthamo wokuhlola kusenesikhathi. Hlola ukuthunyelwa okukhulu okudingekayo futhi vumela amaqembu ahumushe kahle imiphumela.

7

Ukuphathwa kwe-ML eguqukile kungakuqonda kabi ama-metric, imiphumela yokuhlola engafanele, futhi yenze amaphutha amasu. Kubaluleke ngokulinganayo ukufundisa abenzi bezinqumo njengoba kufundisa amaqembu e-Engineering.

Ngake ngasebenza neqembu elalinabo bonke ubuchwepheshe ababeludingayo, kanye nama-mlops aqinile kanye nokuhlolwa kwe-A / B kodwa abaphathi bebengazi ukuthi bazisebenzisa kanjani. Babesebenzisa izivivinyo ezingalungile zezibalo, babulala izivivinyo ngemuva kosuku olulodwa lapho “ukubaluleka kwezibalo” kutholakele (imvamisa ngokubona okumbalwa kakhulu), kanye nezinto ezethulwe ngaphandle komthelela ongalingani. Umphumela: Ukwethulwa okuningi kwaba nomthelela omubi. Abaphathi ngokwabo kwakungebona abantu ababi, bamane nje abaqondanga ukuthi basebenzisa kanjani amathuluzi abo.

8. Amamethrikhi athunyelwe kabi

Ngenkathi izinhlangano ze-ML / DS zidinga ukuqondiswa kwebhizinisi, lokho akusho ukuthi kudingeka zibe nemvelo yebhizinisi. Abasebenzi be-ML bazohambisana nanoma yini ama-metric abonwa kubo uma bezwa ukuthi banembile. Uma izinhloso ze-ML zifakwe kabi ngemigomo eqinile, khona-ke umphumela uzobe usuphambukiswa. Isibonelo, uma inzuzo yilokho inkampani akufunayo kepha ukukhulisa ukuguqulwa komsebenzisi okusha kuyinjongo yenhlangano ye-ML, bazokhulisa ukukhula okungenakuzuzisa ngokungeza abasebenzisi be-Iyunithi Ababi abangabuyi.

Leli yindawo ebuhlungu yezinkampani eziningi. Inkampani yokulethwa kokudla efisa ukukhula. Abaphathi babona ukuguqulwa okuphansi kwabasebenzisi abasha njengenkinga yokunqanda ibhizinisi kusuka ekukhuleni kwemali. Ithimba le-DS lacelwa ukuthi lixazululwe ngokwenza okuthandwa nguwe kanye nokuphakanyiswa kwamakhasimende. Inkinga yangempela ibigcinwa, abasebenzisi abaguquliwe ababuyanga. Esikhundleni sokugcinwa, iqembu ligxile ekuguqukeni, ligcwalisa amanzi ngempumelelo kwibhakede elivuzayo. Noma izinga lokuguqulwa lithathwe, alihunyushelwa ekukhuleni okuzinzile. Lawa maphutha awekho ibhizinisi noma usayizi wezezimboni – lawa ngamaphutha wendawo yonke.

Zingavinjelwa noneze. I-AI ne-ML yenza umsebenzi lapho wenziwe waklanywa ezimisweni ezizwakalayo, eklanyelwe ukuxazulula izingqinamba zangempela, futhi enziwa ngokucophelela ebhizinisini. Lapho yonke imibandela ilungile, i-AI ne-ML iphenduka ubuchwepheshe obuphazamisayo ngamandla okuthuthukisa wonke amabhizinisi.

Ukudla okudlela endlini: Yenza ama-ML metric avumelane nezinhloso zebhizinisi leqiniso. Ukulwa izimbangela, hhayi izimpawu. Yazisa ukusebenza kwesikhathi eside, hhayi amamethrikhi wesikhathi esifushane.

Ukugcina

Indlela eya e-AI / ML PREMENT ingaphansi kwama-algorithms aphesheya kwegazi nokuningi ngokuvuthwa kwenhlangano. Amaphethini abonakala: Ukwehluleka kuvela ekuphuthumeni kube yinkimbinkimbi, izisusa ezingezinhle zokuziphatha kabi, nokunganaki ingqalasizinda yesekethe. Impumelelo ifuna ukubekezela, isiyalo, nokuvuleleka kokuqala okuncane.

Izindaba ezihle ukuthi wonke la maphutha agweme ngokuphelele. Amafemu abeka ingqalasizinda yedatha endaweni yokuqala, gcina ukuxhumana okusondele phakathi kwamaqembu ezobuchwepheshe nawamabhizinisi, futhi awaphazanyiswa yi-FADS azothola ukuthi i-AI / ML ithola ukuthi i-AI / ML yenza ngokuqondile lokho okuthembisayo. Ubuchwepheshe busebenza, kepha kufanele kube sezingeni eziqinile.

Uma kukhona i-tenet eyodwa ebopha konke lokhu ndawonye, ​​yile: AI / ML iyithuluzi, hhayi lapho uya khona. Qala ngenkinga, qinisekisa isidingo, ukuthuthukisa ngokuqondile, futhi ulinganise njalo. Lawo mabhizinisi asondela kulo nale mindset awagcini nje kuphela. Esikhundleni salokho, badala umehluko wesikhathi eside wokuncintisana ohlanganisa isikhathi.

Ikusasa akuyona eyamafemu ngamamodeli amasha, kodwa kuma-Firms anesiyalo sokuwasebenzisa ngokunengqondo.