使用 SLM 和微型代理构建高精度 AI 系统.pdf

1、Composing High-Accuracy AI Systems with SLMs and Mini-AgentsSharon ZhouJune 10,2025IntroSharon Zhou,PhD something new:)Founder&CEO,Lamini Stanford CS Faculty in Generative AI Stanford CS PhD in Generative AI(Andrew Ng)MIT Technology Review 35 Under 35,foraward-winning research in generative AI Creat

2、ed largest Coursera courses(Generative AI)Google Product Manager Harvard Classics&CSScaling AI Agents with an Open EcosystemSharon ZhouJune 10,2025IntroHigh-accuracy AI systemsFactual accuracy is hard for LLMs,especially SLMs and mini-agentsHow do you successfully make SLMs and mini-agents useful fo

3、r you?How do you scale data for them?Best practices&learnings across the enterpriseScaling AI agents and AI progress with an open ecosystemCompute-bound AI&Scaling LawsStack layersPlayers,status,and future outlooksOpen ecosystemQ&AAgendaComposing High-Accuracy AI Systems with SLMs and Mini-AgentsSha

4、ron ZhouJune 10,202533%of enterprise apps ACCORDING TO GARTNERwill be powered by generative AI and agentic AI by 2028General LLMs:pretty good at everything,but perfect at nothingThe LLM doesnt know a fact that is similar,is actually wrong Pretrained LLMs rely on similarities.UntrainedPretrained,Fine

5、tuned,RAGThe next frontier of high-value use cases need factual reasoningBest for promptingBest for RAGBest for advanced RAG and fine-tuningText-to-SQLCustom codebaseCross-document reasoningLarge catalog product recommendationsFunction callingBasic information retrievalGeneral content creationBasic

6、Q&AText summarizationBasic assistantsAgentic workflowsGetting your facts right in the data is important.The model will treat them as The Truth.To make(factual)learning algorithms work,you need these other ingredients:HIGH-QUALITY TRAINING DATAMaking sure your evaluation is objective,not subjective,g