Waxaan ka soo bandhigiyaan Dynamic Product Ads ee Twitter, oo aan ka soo bandhigi karaa million users to hundred million e-commerce products in real time. The output was the top 5-6 products that each user was most likely to buy. The system used product and user embeddings with classic ML models to serve personalized ads at Twitter's scale. We saw 15-18% improvements in CTR and 12% improvements in conversion rates compared to brand ads. Sida loo yaabaa in ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ah mid ka mid ah mid ka mid ah mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ah mid ka mid ah mid ah mid ka mid ah mid ah mid ka mid ah mid ka mid ah mid ah mid ka mid ah mid ah mid ka mid ah mid ah mid ka mid ah. Sida loo isticmaali karaa LLM ee ka mid ah 20% ee nidaamka, si ay u isticmaali karaa in ay u isticmaali karaa in ay u isticmaali karaa in ay u isticmaali karaa in ay u isticmaali karaa. Qalabka Orixinal Sida loo yaabaa, waxaa laga yaabaa in ay ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah Qalabka Products Waxaan isticmaali karaa metadata ee cadaadiga, sida title, description, category, price, iwm, iyo codeeyo in 128-dimensional vector space. Waxaan sidoo kale loo isticmaali karaa si ay u isticmaali karaa user engagement in this product or conversion patterns to calculate embeddings. Shuruudaha user Users waxay ku yaalaa by vector-ka ku salaysan si ay u dhismaha oo ku saabsan adeegga ah ee platform, info profile, iyo iibsado warshadaha. Qalabka Matching Waayo, waxaan isticmaali karaa mid ka mid ah mid ka mid ah mid ka mid ah mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ah mid ka mid ah mid ah mid ka mid ah mid ah mid ka mid ah mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ah mid ka mid ah mid ka mid ah mid ah mid ka mid ah mid ka mid ah mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ah mid ka mid ah mid ka mid ah mid ah mid ka mid ah mid ah mid ka mid ah mid ka mid ah mid ah mid ka mid ah mid ah mid ka mid ah mid ah mid ka mid ah mid ah mid ka mid Model iyo ANN (waxariisada ugu soo saarka ah) waxaa laga yaabaa, waxaa laga yaabaa, oo ah oo ugu weyn, ka mid ah si ay u hesho Twitter. Sidee waxa uu ka soo xiriir in la aasaasay? Haddii aad u sameeyaan nidaamka this today, this is what I’d actually change. Qalabka wax soo saarka ugu fiican ee coders LLM Markaad ugu badan waa in la soo saarka wax soo saarka wax soo saarka wax soo saarka wax soo saarka wax soo saarka wax soo saarka wax soo saarka wax soo saarka wax soo saarka wax soo saarka wax soo saarka wax soo saarka wax soo saarka wax soo saarka wax soo saarka wax soo saarka wax soo saarka wax soo saarka wax soo saarka wax soo saarka wax soo saarka wax soo saarka wax soo saarka wax soo saarka wax soo saarka wax soo saarka wax soo saarka wax soo saarka wax soo saarka wax soo saarka wax soo saarka wax soo saarka wax soo saarka wax saarka wax soo saarka wax soo saarka wax saarka wax soo saarka wax saarka wax soo saarka wax saarka wax soo saarka wax saarka wax saarka wax soo saarka wax saarka wax soo saarka wax saarka wax saarka wax Sida loo yaqaan "Shoes Running" waxaa loo yaqaan "sneakers for jogging" oo ka mid ah "sneakers for jogging" oo ka mid ah in ay ku yaqaan "sneakers for jogging" oo ka mid ah "sneakers for jogging" oo ka mid ah "sneakers for jogging" oo ka mid ah "sneakers for jogging" iyo "sneakers for jogging". Waayo, sidoo kale waxaa laga yaabaa in la sameeyo. all-MiniLM-L6-v2 Waayo, waxa uu ka mid ah wax soo saarka Nike oo ka mid ah 'Air Max 270 React' waxay ka mid ah in ay ka mid ah in ay ka mid ah in ay ka mid ah isticmaalka oo ka mid ah 'cushioned running shoes' ama 'athletic sneakers' oo ka mid ah in ay ka mid ah wax soo saarka keyword. Qalabka Cold-Start LLM ayaa sidoo kale loo isticmaali karaa in ay ka mid ah wax soo saarka soo saarka ah. Markaad ka mid ah wax soo saarka cusub waa in la soo saarka, LLM waa in la soo saarka si ay u soo saarka wax soo saarka, wax soo saarka, wax soo saarka, wax soo saarka, wax soo saarka, wax soo saarka, wax soo saarka iyo wax soo saarka. Sida loo yaqaan 'LLMs' waxay ku yaalaa arka ah? Qalabka Hybrid Waxaan sidoo kale loo isticmaali karaa ML classic for storing and serving layers. The architecture would look like: 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 CODE CODE CODE CODE CODE CODE CODE CODE CODE CODE CODE QEEBE Shuruudaha ugu soo saarka ah ee loo soo saarka Products Candidate per user Haku Qalabka iyo Ranking Final Haku Markaas ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah. LLMs waxay ku raaxeyn karaa waxyaabaha ee million users in less than 10 milliseconds. Waxay ku yaalaa in la xira si ay u xiran karo macluumaadka dhismaha iyo si ay u xiran tahay si ay u xiran tahay. model classical waa in macluumaadka microseconds. At Twitter's scale, diference between 1ms and 10ms inference time translates in million dollars in infrastructure costs and measurable drops in user engagement. Qalabka dhismaha iyo dhismaha iyo dhismaha, dhismaha iyo dhismaha, dhismaha iyo dhismaha, dhismaha iyo dhismaha, dhismaha iyo dhismaha, dhismaha iyo dhismaha, dhismaha iyo dhismaha, dhismaha iyo dhismaha, dhismaha iyo dhismaha, dhismaha iyo dhismaha, dhismaha iyo dhismaha, dhismaha iyo dhismaha. Sida loo isticmaalo ad copy? Waxaa laga yaabaa in ka badan in internet oo loo isticmaali karaa LLMs si ay u soo saarka ad-hoos ah in la furan ama si ay u raaxaysaa oo aad isticmaali karaa in real-time. this is where it gets hard to decide if LLMs would actually be useful or not. Shuruudaha ugu horeysay ee shuruudaha ugu horeysay ee shuruudaha ugu horeysay ee shuruudaha ugu horeysay ee shuruudaha ugu horeysay ee shuruudaha ugu horeysay ee shuruudaha ugu horeysay ee shuruudaha ugu horeysay ee shuruudaha ugu horeysay ee shuruudaha ugu horeysay ee shuruudaha ugu horeysay ee shuruudaha ugu horeysay ee shuruudaha ugu horeysay ee shuruudaha ugu horeysay ee shuruudaha ugu horeysay ee shuruudaha ugu horeysay ee shuruudaha ugu horeysay ee shuruudaha ugu horeysay ee shuruudaha ugu horeysay ee shuruudaha ugu horeysay ee shuruudaha ugu horeysay ee shuruudaha ugu horeysay ee shuruudaha ugu horeysay ee shuruudaha ugu horeysay ee shuruudaha ugu hore Nidaamka waxaa laga yaabaa? Sida loo isticmaali karaa in ay u isticmaali karaa in ay u isticmaali karaa in ay u isticmaali karaa in ay u isticmaali karaa in ay isticmaali karaa in ay isticmaali karaa in ay isticmaali karaa in ay isticmaali karaa in ay isticmaali karaa in ay isticmaali karaa in ay isticmaali karaa in ay isticmaali karaa in ay isticmaali karaa in ay isticmaali karaa in ay isticmaali karaa in ay isticmaali karaa in ay isticmaali karaa in ay isticmaali karaa in ay isticmaali karaa in ay isticmaali karaa in ay isticmaali karaa in ay isticmaali karaa in ay isticmaali karaa in ay isticmaali karaa in ay isticmaali karaa in ay isticmaali karaa in ay isticmaali karaa in ay isticmaali karaa in ay isticmaali karaa in ay isticmaali karaa in ay isticmaali karaa in ay isticmaali karaa in ay isticmaali karaa. Understanding user intent Waayo, sidoo kale waxaa laga yaabaa in ka mid ah mid ka mid ah mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ka mid ah. Latency requirements Qalabka ugu horeysay ee loo yaqaan "Cold Start for New Products or Users" iyo "data quality issues when catalogues have missing or incorrect product descriptions" waxaa laga yaqaan "Cold Start for New Products or Users" iyo "Data Quality problems when catalogues have missing or incorrect product descriptions" waxaa laga yaabaa. Last-mile problems Dhismaha dhismaha dhismaha dhismaha dhismaha dhismaha dhismaha dhismaha dhismaha dhismaha dhismaha dhismaha dhismaha dhismaha dhismaha dhismaha dhismaha dhismaha dhismaha dhismaha dhismaha dhismaha dhismaha dhismaha dhismaha dhismaha dhismaha dhismaha dhismaha dhismaha dhismaha dhismaha dhismaha dhismaha dhismaha dhismaha dhismaha dhismaha Iteration speed Qiyaasta ah waa: Waayo Bottleneck waa? Markaas oo ka mid ah mid ka mid ah "how do you would build it today with modern AI" waxay ka mid ah ka mid ah mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ah mid ka mid ah mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ah mid ka mid ah mid ka mid ah mid ah mid ka mid ah mid ah mid ka mid ah mid ah mid ka mid ah mid ah mid ka mid ah mid ah mid ka mid ah mid ah mid ka mid ah mid ah mid ka mid ah mid ah mid ka mid ah mid ah mid ka mid ah mid ah mid ka mid ah mid ah mid ka mid ah mid ah mid ah mid ka mid ah mid ah mid ka mid ah mid ah mid ah mid Waayo, dhismaha dhismaha ayaa ka mid ah cilmi-baarista. Waxaa la mid ah in ay u baahan tahay in ay ka mid ah dhismaha macaamiisha, si ay u baahan tahay in ay ka mid ah cilmi-baarista data, iyo si ay u mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ka mid ah mid ah mid ka mid ah mid ah mid ka mid ah mid ah mid ka mid ah mid ah mid ka mid ah mid ah mid ka mid ah mid ah mid ka mid ah mid ah mid ka mid ah mid ah mid ka mid ah mid ah mid ka mid ah mid ah mid ah mid ka mid ah mid ah mid ka mid ah mid ah mid ka mid ah mid ah mid ah mid ka mid ah mid ah mid ah mid ka mid ah mid ah mid Haddii aad u baabuurta this nidaamka hadda, waxaan u dhigi doonaa 20% oo aan la "sano LLM si loo isticmaalo in la soo saarka ugu fiican" iyo 80% on la mid ah macluumaadka, wax soo saarka data, wax soo saarka, iyo kharashka macaamiisha. Qalabka Unpopular Shirkadda technology waa la soo bandhigay in la xigtay. Waxaa ka mid ah hype oo ka mid ah blockchain iyo Web3 ka mid ah ka mid ah ka mid ah ka mid ah ka mid ah ka mid ah ka mid ah ka mid ah ka mid ah ka mid ah ka mid ah ka mid ah ka mid ah ka mid ah ka mid ah ka mid ah ka mid ah ka mid ah ka mid ah ka mid ah.
