What's next for AI agentic workflows ft. Andrew Ng of AI Fund

スライドのタイトルは「Agentic Reasoning」

LLM-based agents

Non-agentic workflow (zero-shot)

LLMは修正できない（その条件でもうまくやるが）

Agentic workflow

Write an essay outline on topic X

Do you need any web research?

Write a first draft

Consider what parts need revision or more research

Revise your draft

HumanEval

GPT-3.5 + Agentic workflowでGPT-4 + Non-agentic workflowを上回る

Agentic Reasoning Design Patterns（まとめ 11:10）

上2つは2024/03時点でrobust、下2つはemerging

Reflection

人: Please write code for (task)

LLM自身に書いたコードを読んで指摘させる

Check the code carefully for correctness, style and efficiency, and give constructive criticism for how to improve it. (4:46)

（IMO：生成AIコーディングだとここをlinter (tool)が担うかも）

Code AgentとCritic Agentでやる例

参考

Self-Refine

Reflexion

Tool use

例：Web search、Code execution

Computer visionかららしい

参考

Gorilla

MM-REACT

Planning

HuggingGPTの例

Please generate an image where a girl is reading a book, and her pose is the same as the boy in the image example.jpg, then please describe the new image with your voice.

Pose Determination

Pose-to-Image

Image-to-Text

Text-to-Speech

（それぞれでモデルを見つけてくる）

参考

Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

HuggingGPT: Solving AI Tasks with ChatGPT and its Friends in Hugging Face

Multi-agent collaborations

ChatDev

複数のLLMによるdebateで性能が引き上がる

参考

Communicative Agents for Software Development

AutoGen

Reviseさせるのでトークン生成の速さが重要

IMO：GPT-3.5がAgentic workflowでGPT-4を上回ったのが示唆的

https://www.youtube.com/watch?v=sal78ACtGTc