引领未来：推进人工智能代理的基础设施建设.pdf

1、ADVANCING INFRASTRUCTURE FORAgentic AIPIONEERING THE FUTURESHRUTI BHATHow do you go from data to intelligentapps agents?Credit:Federal Reserve Bank StLiHow doesGenAIadoptioncompare withthat of othertechnologiestwo years in?Zero shotMulti-stepvsAgentic-ness lies on a spectrumPlan and executemultiple

2、LLM calls toachieve an outcomeMulti-stepprocessingVoice&image inputsand tools like search,email,calendar Planning and tool useLong/short termmemory,and learningfrom outcomes Memory andreflectionCollaborate withother agents toorchestrate Multi-agentcollaborationCredit:DeepLearning.AICredit:Andrew NgG

3、PT 3.5 withagenticworkflowoutperformsGPT 4Where are we headed?Evolution and landscape of Agentic AIActCollaborateLearnAutonomy totake actionsCollaborate withother agentsReflect and learnfrom actionsEvolution Not scalable.Wide scope.Entire roles and workflowsEarlier:Great demos Narrow scope,augmentsp

4、ecific tasks in a workflowToday:Specific tasks Planners,workers,critics,supervisors.Your own personalset of minions Future:Multi-agent 123Credit:McKinsey Digital Sierra AIEma UnlimitedSifthubResolve.aiStartup StoriesSalesforceSnowflakeDatabricksOpenAIEnterprise PlatformsAgentic AI LandscapeOff-the-s

5、helf vs custom enterprise agents Custom AIagentsInfra stack and abstractionsInfrastackDataLayerCoreLLMOrchestration&GovernanceSDKs&APIsContext&MemoryReasoningTrust&securityTools&accessContext:Prompt,RAG,Tune or Train?Create abstractions Reference:Agentforce,SalesforceChallengesandopportunities“Softw

6、are is eating the world but AI isgoing to eat software.”Jensen HuangAccuracyCostLatencyHandling Trust Issues+Observability+Governance+SecurityTransparencyScalability ConcernsModular approach Unknown Unknowns70%20%10%People and processTechno