(Full video)

Preface

This three-hour interview with Ji Yichao, the chief scientist of Manus (later acquired by Meta), is a must-watch for many in the AI industry.

I will gradually put the great things I see in the comments. Let me quickly talk about the areas I find worth paying attention to and sharing.

As a serial AI entrepreneur, Ji Yichao shares his experience in the AI field over the past decade. From Tokenize to LSTM to Transformer-related applications. To the experience of making AI browsers twice.

Then he talks about why Manus succeeded, including how they solved the problems that Cloud Providers and Model Providers couldn’t - a good toolkit that allows LLMs to automatically plan things.

He also mentions that AI Agents are very similar to manufacturing, because there are so many parts that need to be optimized. Regarding Manus’ product planning and direction, he also believes that the right thing to do is: "Decide what not to do".

There are many technical concepts that are quickly mentioned, but they are actually very deep. I will write them down one by one in the comments.

Regarding MCP Usage

Quick link https://youtu.be/UqMtkgQe-kI?t=10342

Regarding MCP usage, Manus is relatively conservative, because this dynamic discovery tool method of MCP will pollute the Action Space, which will reduce the cache hit rate. The reduced cache hit rate will cause the cost to skyrocket. Improvement method:

• MCP call method not in the native Action Space

He also said that this was written as a blog by Anthorpic, here:Code execution with MCP: Building more efficient agents how to use a code execution environment to improve the efficiency of using MCP (Model Context Protocol) to connect AI agents with external systems. MCP is an open standard designed to solve the problem of connecting AI agents with tools and data. The article points out that with the widespread application of MCP, direct tool calls will lead to context window overload and intermediate results consuming too many tokens. By presenting the MCP server as a code API, agents can manage context more efficiently, reduce token usage, and improve efficiency. Important points

• Challenges of MCP:
• Tool definition overloads the context window.
• Intermediate tool results consume extra tokens.
• Advantages of code execution:
• Agents can load tools on demand and process data in the execution environment.
• Reduce token usage, reduce costs and delays.
• Provide benefits of privacy protection and state management.
• Implementation of code execution:
• Use TypeScript to generate a file tree of available tools.
• Agents explore tools through the file system and only load the required definitions.
• Other benefits of code execution:
• Perform data filtering and transformation.
• Use familiar code patterns for control flow.
• Operations that protect privacy.
• State persistence and skill saving.

Solutions

• Code execution environment:
• Treat the MCP server as a code API.
• Agents run code in the execution environment, reducing the burden on the context window.

Notes

• Security and infrastructure requirements:
• Requires a secure execution environment, appropriate sandboxing, resource limits, and monitoring.
• These infrastructure requirements increase operational overhead and security considerations.

Mentioning OpenAI’s 5-Level Agent

https://cdn.openai.com/business-guides-and-resources/a-practical-guide-to-building-agents.pdf

• Level 1: Conversational AI/Chatbots
• Level 2: Human-Level Problem Solving/Reasoners
• Level 3: Agents
• Level 4: Innovators
• Level 5: Organizers

from https://youtu.be/UqMtkgQe-kI?t=9650

Papers that have influenced AI progress in my mind

When asked which papers have influenced AI progress in your mind:

• FLAN-T5 (Scaling Instruction-Finetuned Language Models)

• Through fine-tuning parameters, the 11B FLAN-T5 also has good results.

• Word2Vec famous paper (Efficient Estimation of Word Representations in Vector Space)

• The Word2vec algorithm, through an efficient mathematical structure, allows computers to convert words into vectors with semantic associations, officially opening the golden age of deep learning in the field of natural language processing.

Discussions related to entrepreneurship

Also related to entrepreneurship: https://youtu.be/UqMtkgQe-kI?t=3783

• When a group of not-so-stupid people are idle, they will have good ideas
• “For every complex problem there is an answer that is clear, simple, and wrong.” - H. L. Mencken