陳天宇-GenAl驅動的xGen電商AI平臺產品實踐與思考.pdf

《陳天宇-GenAl驅動的xGen電商AI平臺產品實踐與思考.pdf》由會員分享，可在線閱讀，更多相關《陳天宇-GenAl驅動的xGen電商AI平臺產品實踐與思考.pdf（31頁珍藏版）》請在三個皮匠報告上搜索。

1、陳天宇eBay 人工智能平臺 產品負責人 GenAI驅動的xGen電商AI平臺 產品實踐與思考 2024 全球產品經理大會PM-SUMMITGenAI-重塑電商業務效能GenAI-重塑開發范式xGen AI平臺產品實踐未來展望GenAI-重塑電商業務效能We believe eBay is best positioned to capture upside from gen AI in 24,to the extent its seller-focused features drive listing velocity and quality.-Nathan Feather(2024/04/1

2、8)https:/ GenerationRecommendationImage GenerationAI AssistantInteraction IntelligenceMarketingText GenerationPersonalized lifestyle,shoppable images,powered by AI,combining items from multiple fashion categories and featuring them in a real-life setting.Features part of MVP Image generation Style-b

3、ased personalization Visual searchCategoriesFashionRecommendationImage GenerationHomepage entry pointShop the look Landing PageShop the look Landing Page on scrollShopping Assistant is actually a Multi-AGENT Bot-Planning-Multi-modal knowledge-ActionAI AssistantInteraction Intelligence另一大賽道-研發效能Veloc

4、ityHow could AI capabilities improve our work efficiencyBusinessHow could AI empower our business and drive GMV/MAU increaseAn internal survey conducted in 2023eBayGenAI-重塑開發范式LLM Solution StackOpsLLM dev lifecycleLLM TrainingLLM ResearcherHuge data(xx T tokens)Large Model Pretrain(xx B params)Cont

5、pretrain/Instruct tuneX monthsModel CatalogIn-house LLMsOpen-source LLMsCloud LLMsResearchApplied ResearcherPrompt EngLLM Fine-TuneRAGand/ordatadataOrchestrationX weeks monthsML EngineerDeployMaintain/UpgradeIntegration TestX daysPlanningActionLLM pretrain 的 門檻&scaling law洞見1Prompt Engineering 取代 Fe

6、ature Engineering成為最大的“trick”洞見2工程&model 的協同性更強(like RAG)-Applied Researcher 需要更強大的 dev tool ML Engineer 需要更多的 數據理解洞見3xGen AI平臺產品實踐回顧-去年此時的eBay AI 平臺計算集群數據湖/數倉數據引擎/管道*業務(domain)MLOps/ML 分析代碼倉庫/鏡像倉庫/CICD人工智能平臺特征工程平臺訓練平臺推理平臺特征倉庫(Feature Store)實驗模板模型倉庫Jupyter NotebookWeb PortalSDKAI 資產生命周期構建/發布用戶接口下一代

7、eBay AI 平臺的產品矩陣 計算集群數據湖/數倉數據引擎/管道*業務(domain)MLOps/ML 分析代碼倉庫/鏡像倉庫/CICD人工智能平臺特征工程平臺訓練平臺推理平臺特征倉庫(Feature Store)實驗模板模型倉庫Jupyter NotebookWeb PortalSDKAI 資產生命周期構建/發布用戶接口VectorDBKnowledge RepoGenAI SandboxModular PipelineProductivity 1-Prompt StudioPrompt商業大模型開源模型自研模型gpt 3.5/4/4oSmall SizeSmall SizeGemini

8、xxMedium SizeMedium SizeClaude xxLarge SizeLarge Size模型選擇Prompt 設計&迭代Leaderboard推理質量Human JudgeAI Judge推理成本ObjectiveModel Selection in 1 dayDatasetProductivity 2-Lora fine-tune StudioObjectiveLora FT in 1 dayHowZero Code1k-3k dataset to startsimplified evaluation approachmanaged LLM&GPUProductivity

9、3-Drag-n-Drop modular pipelineObjectiveML Pipeline in 1 weekHowLow CodePre-built modulesLLM/CV modelsStorage(VectorDB,KG)Batch/NRT inference/RAGProductivity 4-借助 Ray.io 重塑 ML infraRay DataRay TuneRay CoreRay TrainRay ServeRay Tech StacksRay ClusterRay JobRay ServiceTess(Kuberay)23-Greg Brockman,Pres

10、ident of OpenAIProductivity 4-借助 Ray.io 重塑 ML infraML PlatformUnified Feature StoreTraining PlatformUnified Inference Platform(UIP)MLP Control PlaneAI Hub-Drag-n-Drop pipelineNotebooks&SDKsExperimentation managementRay DataRay TuneRay CoreRay TrainRay ServeRay Tech StacksRay ClusterRay JobRay Servic

11、eTess(Kuberay)產品化Ray.io 的key value5x研發效能提升 in大規模 模型訓練/推理3xGPU Util 提升模型 throughput 提升Responsible AIin E2E model dev lifecycleTraining datasetsScale training and fine tuning Model exploration:coding and trainingModel catalogModel catalogData loading and preprocessing pipelinesRollout to ProductionOnlineEndpoint1.Data Governance2.Model Security Scan3.Resonposible AI Audit4.content modernization0.ML asset lineage in the E2E dev flow未來展望垂域多尺寸“小”模型+Synergistic Agents未來展望-1Prompt Eng將被逐步取代LLM開始具備慢慢思思考考能力未來展望-2行業 LLM 的迭代速度取決于 feedback loop的完善程度未來展望-3THANKS