利用基于LLM的AI代理实现更高层次的自主性.pdf

1、Reach Next-level Autonomy with LLM-based AgentsTingyi LiEnterprise Solutions ArchitectAmazon Web ServicesEnterprise Solutions ArchitectThought leader in AIPublic SpeakerPart-time illustratorTingyi LiWriting novels,MusicFinance,SwimmingLimitations of Large Language Models(LLMs)that need to be overcom

2、e1.No easy way to incorporate latest data or specific data inside the modelWeb searches/Information retrieval2.AI models are incredibly smart but shockingly stupid Constitutional AI,Guardrails3.Models work as black box.Cant access external systems,needs engineeringLangchain Agents 4.Doesnt have shor

3、t/long-term memoryConversational memoryUser experienceCompletions are expensive!Engineering vs TechnologyWhat takes for LLM-based applications to become LLM-based Agents?The Rise and Potential of LLM-Based Agents:A Survey.AutonomyAutonomyReactivenessReactivenessProactivenessProactivenessLLM Powered

4、Autonomous AgentsperceptionMulti-modality PerceptionsOnline Shopping Assistant Agent24Jurassic-2 Ultra Action GroupsAbstractionsContext&State ManagementPlanning&ReflectionPlanning&Reflectionhttps:/ for Amazon BedrockE X T E N D T H E F U N C T I O N A L I T Y O F F O U N D A T I O N M O D E L S T O

5、T A K E A C T I O N O N C O M P A N Y S Y S T E M S A N D D A T ACOMPLETES TASKREQUESTAMAZONBEDROCKAGENTFMGenerate a sequence of stepsFM EXECUTES PLAN123Information neededAPIs to call and when to call themOrchestrationOrchestrationAgentOneAgentOneAgentTwoAgentTwoAgentThreeAgentThree.LangGraphchatdev

6、What are the challenges?What are the challenges?AgentOneAgentOneAgentTwoAgentTwoAgentThreeAgentThreeFailed planning for complex tasksFailed to choose/use toolsLost in the middleFailed ROIHow to improve?Prompting makes a huge difference.Finetuning matters.When to finetuneFacts&KnowledgeStyle of Answe