[TIL] A Three-Hour Interview with Ji Yichao, Chief Scientist at Manus (Acquired by Meta)

Source: Dev.to

January 5th, 2026
Full Video

Overview

• Ji Yichao reflects on a decade of building AI startups, covering the evolution from Tokenize → LSTM → Transformer applications and his two experiences building AI browsers.
• He explains why Manus succeeded by solving problems that cloud providers and model providers could not—specifically, a robust tool library that enables LLMs to autonomously plan tasks.
• AI agents are likened to manufacturing processes due to the extensive optimization required.
• Manus’s product strategy emphasizes “deciding what not to do.”

Quick link to the interview segment:

Manus & Model Context Protocol (MCP)

Manus adopts a conservative approach to MCP usage because dynamic tool discovery can pollute the Action Space, reducing cache‑hit rates and increasing costs. The proposed improvement is to invoke MCP outside the native Action Space.

MCP Challenges

• Tool definitions overload the context window.
• Intermediate tool results consume extra tokens.

Advantages of Code Execution

• Agents can load tools on demand and process data within an execution environment.
• Reduces token usage → lower costs and latency.
• Improves privacy protection and state management.

Implementation Details

• Uses TypeScript to generate a file‑tree of available tools.
• Agents explore the file system, loading only the necessary definitions.
• Enables data filtering, transformation, and privacy‑preserving operations.
• Supports state persistence and skill saving.

Security & Infrastructure

• Requires a secure sandboxed environment with resource limits and monitoring.
• These operational requirements add overhead and security considerations.

Reference: “Code execution with MCP: Building more efficient agents” (Anthropic blog).
PDF on building agents (OpenAI)

Levels of AI Capability

1. Conversational AI / Chatbots
2. Human‑Level Problem Solving / Reasoners
3. Agents
4. Innovators
5. Organizers

(See timestamp: )

Influential Papers Mentioned

• FLAN‑T5 – Scaling Instruction‑Finetuned Language Models

  • Achieved strong results with an 11B FLAN‑T5 model.
  • arXiv:2210.11416

• Word2Vec – Efficient Estimation of Word Representations in Vector Space

  • Introduced efficient mathematical architecture for converting text into semantically related vectors, ushering in the deep‑learning era for NLP.
  • arXiv:1301.3781

Entrepreneurship Insight

• “When a group of not‑so‑dumb people have nothing to do, great ideas emerge.” (timestamp: )
• Quote: “For every complex problem there is an answer that is clear, simple, and wrong.” – H. L. Mencken