Ndingathanda i-Dynamic Product Ads kwi-Twitter, apho sinxibelelanisa iimveliso zeemigodi zeemigodi zeemigodi zeemigodi zeemigodi zeemigodi zeemigodi zeemigodi zeemigodi zeemigodi zeemigodi zeemigodi zeemigodi zeemigodi zeemigodi zeemigodi zeemigodi zeemigodi zeemigodi zeemigodi zeemigodi zeemigodi zeemigodi zeemigodi zeemigodi zeemigodi zeemigodi zeemigodi zeemigodi zeemigodi zeemigodi zeemigodi zeemigodi zeemigodi zeemigodi zeemigodi zeemigodi zeemigodi zeemigodi zeemigodi zeemigodi. Yintoni iinyanga ezininzi ezidlulileyo. Kodwa ngoku, wonke umntu uqhagamshelane nge-AI kunye ne-Big Language Models njengoko izixazululo yonke into. Ngoko ke, ndingathanda ukuba ngaba ndiyifake i-Dynamic Product Ads namhlanje, ngaba ndiyisebenzisa i-LLMs? Kwaye engaphezulu kakhulu, yintoni engabonakali, kwaye ngoko ke? Ukusabela: Ngaba usebenzisa i-LLMs malunga ne-20% ye-system, ikakhulukazi ukuguqulwa kwe-embeddings, kwaye ukugcina yonke into elinye. Umgangatho yethu yokuqala Iingxaki elandelayo: ukunxibelelana iimveliso kwiinkonzo ezininzi ukuba bafumane kunye nokupakisha, kwaye zithembisa ngokukhawuleza ixesha yabo. Kukho iimilioni zeemveliso ezahlukileyo kunye neengcali. Ngaphezu kwalokho, kukho iimilioni zeemveliso zithembisa ngexesha elifanayo. I-system kufuneka ifake izibuyekezo zeemilioni zeemveliso kwi-sub-millisecond latency. I-Ad Serving pipeline kufuneka ifumaneke ukuhlaziywa kwizinga ze-50 milliseconds, ngexesha elide. Umgangatho yendaba kakhulu (kwi-2022 okungenani). Iimveliso ze-embeddings Ukusebenzisa i-metadata yeenkcukacha zayo, njenge-title, i-description, i-category, i-price, njl, kwaye i-coded kwi-128-dimensional ye-vector space. Ukusetyenzisa i-embeddings. Iimveliso ze-user embeddings Abasebenzisi wabanjwa nge-vector ezisekelwe kwi-signals efana neengxaki zayo kwi-platform, iinkcukacha ze-profile, kunye neengxaki ezidlulileyo. Kwakhona i-geographics kunye nexesha lokugqibela ziye zihlanganisa apha. Iimodeli ye-Matching Kwixesha lokugqibela, siye usebenzisa indlela yeetaphu ezimbini. Okokuqala, siya kuqhuba ukufuna ngokushesha kwizixeko zangaphakathi zangaphakathi ukufumana iimveliso ze-candidate ezinxulumene ne-user embeddings. Emva koko, siye usebenzisa i-gradient-boosted decision tree ukufumana amacandelo zangaphakathi, zihlanganisa iimpawu ezongezelelweyo ezifana nexesha elidlulileyo, iingcingo zexabiso, kunye ne-context ezifana nexesha lokugqibela. I-model kunye ne-ANN (i-approximate nearest neighbor) ziye zibonakalayo, zibonakalayo, kwaye okungenani kakhulu, ngokukhawuleza kakhulu kwi-scale yeTwitter. Yintoni ndingathanda kuyo namhlanje? I-2026 ngoku. Ukuba ndithuba le nkqubo namhlanje, apha nto ngathi ndithuba. Imveliso engcono kwi-encoders ye-LLM Ukuphucula okuphumelela kakhulu kuya kuvelisa iimveliso ezilungileyo. Iimodeli ezininzi ezintsha ziyafumaneka kakhulu ukufumana umzekelo kunye ne-context semantic. Ngaphandle kokuhlanganisa iimveliso zeemveliso (iintlobo ezininzi ezininzi ezininzi ziyiqala), ndisebenzisa i-encoder esekelwe kwi-LLM ukuvelisa iimveliso zeemveliso. Yinto kubalulekile ngenxa yokuba iimveliso ebizwa ngokuthi "iingubo zokuhamba" iya kuba ngokufanelekileyo kwi "iingubo zokuhamba", nangona baye zihlanganisa iingxowa zangaphakathi. Iingxowa zangaphambili ze-Hugging Face, ezifana Ukusebenza oku ngempumelelo. all-MiniLM-L6-v2 Xa ngexesha yaba ingxelo ye-Nike ye-catalogue ebizwa ngokuba 'Air Max 270 React' ukuba ama-embeddings zethu ze-2022 awukwazi ukuxhaswa kubasebenzisi abalandeli 'i-cushioned running shoes' okanye 'i-athletic sneakers' ngenxa yokuba akukho ukuxhaswa kweengxelo ze-keyword. Le mveliso yaba 35-40% imibuzo engaphezulu kunokuba iimveliso efanayo kwiiveki yayo yokuqala ngexesha lokufumana idatha epheleleyo. Umdlali esekelwe kwi-LLM uyaziqhelekanga ukuxhaswa kwe-semantic ngokushesha. Ukuphuculwa kweCold-Start Handling I-LLM yenza ukuqhuba kwe-cold-start ngakumbi kakuhle. Xa i-product entsha ifumaneka kwi-catalogue, i-LLM inokufumana iisignals ezininzi kwi-descriptions ye-product, i-reviews, kunye neebhiziyo ukuvelisa i-incubation yokuqala elifanelekileyo. Ngokusho, kubasebenzisi ezintsha kunye neengcali ye-engcubation, ama-encoders ezintsha ziyafumaneka ngakumbi iinkcukacha zayo ze-profile kunye ne-tweets yokuqala (kuya kuba) ukuvelisa iinkcukacha ezininzi. I-Cold-start yaba ngexesha elidlulileyo kwizixazululo zethu ze-user-product matching ezisekelwe kwi-incubations ezincinci. Ngoko ke, apho i-LLM ziyafumaneka kwi-architecture yokwenene? Umgangatho we-hybrid Ndingathanda ukusetyenziswa kwe-ML ye-classical yokuhlanza kunye nokuhlanza i-layers. I-architecture iya kuba: Feature LLM or Classic LLM-based encoder to generate user and product embeddings LLM Match embeddings to generate candidate products per user Classic Final scoring and ranking Classic I-LLM-based encoder yokwenza i-user kunye ne-product embeddings iimveliso Ukudibanisa i-Match Embeddings ukuvelisa iimveliso ze-Candidate per user Ukucinga Ukuhlaziywa kokuqala kunye ne-ranking Ukucinga Yintoni ndiza usebenzisa iimodeli ze-classic yokubala? Iimpawu ziquka i-latency, i-cost, kunye ne-explainability. I-LLMs ayikwazi ukuhlaziywa iimveliso yeemilioni ezisetyenziswa kwizithuba ze-10 milliseconds. Zifumaneka iimveliso ze-overkill yokuxhuma iimfuneko ze-numerical kunye nokwenza ukhetho yokufaka. Iimodeli ze-classic bakwenza oku kwizithuba ze-microseconds. Kwi-scale ye-Twitter, i-difference phakathi kwe-1ms kunye ne-10ms yexesha le- inference ibonisa kwiimilioni zeemali ze-infrastructure kunye neengxaki ezininzi ze-user engagement. Ukusebenza i-LLM ye-inference ye-prediction yeenkcukacha ziye zithengise i-50-100x ngaphezu kwindlela yethu ye-classical. Yintoni malunga ne-Ad Copy? Kukho kakhulu hype kwi-intanethi malunga ne-LLMs ukusetyenziswa ukuvelisa i-ad-copy eyenzelwe kwi-fly okanye ukucacisa malunga neengxaki yabasebenzisi ngexesha elifanelekileyo. Kule nto leyo kuthetha ukuba kubalulekile ukuba i-LLMs iya kuba lula okanye akukho. Ukuguqulwa kwe-ad copy nge-LLMs ibonelela iingxaki ezinzima ezifana ne-hallucinations malunga neengxaki zeemveliso, i-branding ezincinane, kunye neengxaki ze-review kwi-scale. I-system kufuneka ibonisa iimiliyoni ze-ad variations ngosuku, kwaye akukho indlela yokubuyisa kubo ngenxa ye-accuracy kunye ne-brand safety. Umzekelo we-hallucinated malunga ne-product efana ne-"waterproof" xa ayikho, okanye i-"FDA-approved" xa ayikho, uya kuvelisa inkxaso lomgangatho. I-risk ayifunayo ukuchithwa kweengxaki. Yintoni akuyona? Yintoni inkqubo isetyenziswa kwi-2022 okanye kwi-2026 i-LLMs, ingxaki esisodwa: ukufikelela into efunyenwe kwi-signals emangalisayo. Umntu owayenza kwi-tweets malunga neengoma ze-running ingaba i-marathon runner yokuthengisa kwi-pair yayo elandelayo, okanye umbhali omnqweno olungapheliyo owaza umzobo kwi-television. Le ngxaki kufuneka iinkcukacha ezifanelekileyo ukuyisombulula, iinkcukacha zeengcali, kunye neentlobo ezininzi ze-experimentation. Akukho i-architecture ye-model iya kuthetha oku. Understanding user intent i-non-negotiable. Kwi-scale, zonke i-millisecond zihlanganisa. Abasebenzisi bakwazi ukunceda iimvavanyo ezincinane. Iinkqubo ze-ad ayikwazi ukunciphisa ukuxhaswa kwe-timeline. Ndandisa iinkqubo ezininzi ze-infrastructure ze-ML zihlala kwiimvavanyo ze-A / B ngenxa yokongeza i-100ms ye-latency kwi-system eyenziwe kakuhle. Iimodeli ingaba engcono, kodwa isicelo se-latency ibandakanya bonke. Latency requirements Iingxaki efana ne-cold start yeemveliso ezintsha okanye abasebenzisi, kunye neengxaki zekhwalithi yeedatha xa i-catalogu zithintela okanye i-descriptions yeemveliso ezifanelekileyo, ziyafumaneka. Ezi iingxaki ziquka i-architecture ye-model kwaye zinokufuneka ukuyila kwe-system design, yaye akuyona i-architecture ye-model. Last-mile problems Ukuphumelela kwimodeli yokuphumelela. Iqela leyo enokufumana i-10 iimvavanyo ngeveki kunye neemodeli efanelekileyo kunokufumana ngokuqhelekileyo iimvavanyo efanelekileyo ye-1 iimvavanyo ngeveki kunye neemodeli efanelekileyo kakhulu. Ukuphumelela ngokukhawuleza, ukucacisa imiphumela kunye nokuguqula kunokuphumelela kwimodeli kwizinga. Xa siqela i-Dynamic Product Ads, siqela iimvavanyo ze-3-4 ngeveki. Siqela iingxaki ezahlukeneyo ze-embedding, i-ANN algorithms ezahlukeneyo kunye neempawu ezahlukeneyo kwimodeli ye-scoring. Iimvavanyo ezininzi zikhuba. Kodwa iimvavanyo Iteration speed Imibuzo efanelekileyo: Yintoni i-Bottleneck? Ukususela, inkcazelo enkulu ye-"how would you build it today with modern AI" isixeko. Umbuzo kufuneka akuyona into enokwenzeka nge-technology entsha. Kubalulekile, "Yintoni i-bottleneck efanelekileyo kwinkqubo yakho apho i-AI inokunceda?" Kwimeko yethu, i-bottleneck ayikho malunga ne-quality ye-embeddings. I-intent ye-user, ukulawula iingxaki zekhwalithi ye-data kwi-catalogu ye-product, kunye nokulawula iingxaki ze-cold-start, kunye ne-building systems that could handle scale. Ukuba ndandisa le nkqubo namhlanje, ngaba ndandisa i-20% yayo kwi "ukusebenzisa i-LLM ukuvelisa i-embeddings engcono" kunye ne-80% kwiingxaki ezininzi ze-scale, umgangatho weedatha, i-experimentation, kunye nokufumana iingxaki ze-user. Iingcebiso ebandayo I-tech umkhakha ibandakanya iingcebiso ezintsha. Kukho i-hype efanayo malunga ne-blockchain kunye ne-Web3 eminyakeni ezidlulileyo. Kodwa kwakhona, iinkqubo ezininzi zokuvelisa i-ML zokusebenza, zihlanganisa kunye nokwenza imali. Umgangatho we-LLM yenza ukwandisa i-5%, kodwa iya kuba i-10x i-slower kunye ne-100x i-cost. I-AI yeModern ibonakalisa ngokufanelekileyo xa isetyenziswe kwi-bottlenecks, ngaphandle kokubalwa kwinkqubo epheleleyo ebonakalayo.
