Weblink: https://www.crowdcast.io/c/m2ooh0uy9syg

Speak1: Swyx - smol developer

https://docs.google.com/presentation/d/1d5N3YqjSJwhioFT-edmyjxGsPBCMb1uZg0Zs5Ju673k/edit#slide=id.g254e571859c_0_133

Article:

• https://stratechery.com/2018/techs-two-philosophies/
• https://www.latent.space/p/agents
• https://lilianweng.github.io/posts/2023-06-23-agent/

Speaker2: Alex - Agent Eval

https://docs.google.com/presentation/d/1bo5uxaS4JMNt99VmsRdeTFLo9qSIByJiViIVakzF9NQ/edit#slide=id.g22b104eecb9_0_2

• 很難 debug agent failure:
  • failure token
  • CAPCHA

• 三種 Evaluation 方式

• 抓下一堆 dataset

• Alex: https://twitter.com/alexreibman
• Gurkaran: https://twitter.com/aigsingh
• dare: https://twitter.com/dariusemrani
• Jesse Hu https://twitter.com/huyouare

Q&A

1. What is the most affordable (free, local?) LLM for specific Agent Executor / Agent task like decision making, tool selection…?
   • Mpt7b
2. In my experience, the OpenAI functions work really well in deciding what tool(s) to use even in multi-step scenarios. Do you think that a train-of-thought process is used behind the scenes, like ReACT or MLKR? And how useful are they now?
   1. 可以考慮看看 few shot

其他

Agent Hackathon

https://lablab.ai/event/ai-agents-hackathon

AgentEval (第一名)

最後有 OpenAI CEO 演講