俞舟-AI Agent：从实验室到企业.pdf

1、章毅俞舟AI Agents Labs-WorldZhou(Jo)YuColumbia University&Arklex AILLMLLMLLM 2025 Arklex.AI-All rights reservedArklex.AIFounder,CEODr.Zhou(Jo)YuColumbia University CS Professor(CMU PhD)Open-source AI models(1M+downloads)AI Consultant,Microsoft ResearchForbes 30 Under 30amazon prizeWhat are AI agents?Per

2、ception:Multimodal inputs including,text,image,audio,video,touch,etc.Planning(Inner Monologue):Chain-of-Thought reasoning Reflection:meta-reasoning in every stopActions:function/tool calling,embodied actions.2025 Arklex.AI-All rights reservedArklex.AI5The field will continue to push the frontier of

3、capabilityLevel 1:Chatbots(2022 onwards)Level 2:Reasoners(2024 onwards)“Levels of AI”(Sam Altman on Levels of AI)Level 3:Agents(2025 onwards)Level 5:Organizations(20?)Level 4:Innovators(202?)Currently:tasks that take seconds to hours(OpenAI Operator,deep research models,Claude Code,Manus)Eventually:

4、hours to days of workHuge demand for specialists in other fieldsTasks that are at the edge of human performance,or are totally new-New scientific discoveries(Alpha*)-Prize-winning writing-Unsolved mysteriesNext-Gen AI Agents 2025 Arklex.AI-All rights reservedArklex.AIEnterprises are exploring AI emp

5、loyeesAI Agent+Human AgentHuman+SoftwareVS.7Current Agent Ecosystem 2025 Arklex.AI-All rights reservedArklex.AIAI Agent Verticals Applications(Leverage the entire agentic infra to further enhance agents usability through UI/UX,verifiers,human intervention,human teaching,etc to provide real ROIs to e

6、nterprise and individuals)Agent Orchestration Framework(Leverage search,planning algorithms,tool-use,scaffolds,guardrails,to improve agent tasks success-rate,efficiency,security,etc)Agent Foundation Models(Continue training,reinforcement learning to adapt Foundation Models for agent tasks)Foundation