Tsinghua Open-Sources a Batch of "Learning Shrimp": Here Comes a Free "AI Tutor Squad"
The "lobster farming craze" has swept through streets and alleys and become a social language of its own. But for non-technical beginners, how can they learn to use the lobster (OpenClaw) faster and better reap its dividends?
First, let's look at a GIF:
This article includes video content. Please view it in the original WeChat post.
In the GIF above, an "OpenClaw usage tutorial" written by Jia Mu, a well-known KOL in AI applications, is converted with one click into a "Lobster Learning Classroom" — pitched at a beginner's knowledge level, with automatic Q&A interaction and step-by-step guidance, a veritable nanny-level learning guide!
And the behind-the-scenes hero that turns an advanced technical document into a zero-barrier learning classroom in just 10 seconds is OpenMAIC, the world's first multi-agent generative learning AI technology framework, recently open-sourced by Tsinghua University.
Beyond one-click lobster classrooms, OpenMAIC unlocks more solutions for the lobster era — for example, packaging OpenMAIC's capabilities as an MCP Server and Skill, so the lobster can call it and generate courses for learning at any time:
Install the OpenMAIC plugin on the "lobster" and learn autonomously anytime, anywhere.
If in the internet era everyone is a product manager, then in the AI era everyone can be their own classmate, homeroom teacher, even principal (🐶).
Use AI to build a Hogwarts School of Magic, generate a college calculus course at your own pace, and let liberal arts students Vibe Code with ease... On OpenMAIC, countless new ways of learning become possible, breaking the limits of geography, time, and resources.
OpenMAIC's predecessor is "MAIC," a "self-driving" classroom dedicated to using AGI technology.
In early 2024, a discussion on "AI-empowered education" began at Tsinghua, where wild ideas were proposed — could AI agents teach, realizing "AI teaching AI"?
MAIC was born from this and became an important learning platform at Tsinghua. It can teach not only AI knowledge but also various subjects and information far beyond textbooks, and it launched on the National Smart Education Platform with over 20 million cumulative visits.
Not long ago, MAIC's main paper, "From MOOC to MAIC: Reimagine Online Teaching and Learning through LLM-driven Agents," was accepted in JCST's 40th-anniversary special issue hosted by the Chinese Academy of Sciences as the only "AI + education" article, drawing wide academic attention and discussion.
With the MAIC paper accepted by JCST, OpenMAIC, the multi-agent generative learning AI technology framework led by Tsinghua University, has finally been open-sourced to the community for free.
On X, OpenMAIC surpassed 100k views within hours of its open-source release:
➤ GitHub link
🔗 https://github.com/THU-MAIC/OpenMAIC
"AI Teaching AI"
How do humans transmit information?
In the book The Village Teacher, Liu Cixin tells this story:
A carbon-based civilization administers a civilization-level test to Earth, and humanity's fate ultimately hangs on whether a group of children in a remote village have memorized Newton's three laws.
This raises a chilling question: without genetic memory, and with knowledge transmitted only through something as inefficient as sound waves, how does human civilization avoid being locked out by the "Great Filter"?
For thousands of years of human history, the group called "teachers" served as the medium of information transfer between generations. But with the knowledge explosion, class-based instruction by a single group of teachers will find it ever harder to keep up with the transmission of massive knowledge.
The answer may lie in the rapidly advancing AI technology: develop AI into the ultimate carrier of information transfer.
To this end, starting in 2024, the Tsinghua research team threw itself into AI education and launched MAIC, an education agent platform based on large language model technology. It is a "self-driving" AI classroom: from pre-class preparation and in-class interaction to post-class quizzes, AI carries the whole thing, completely changing how knowledge is transmitted.
As soon as MAIC launched, it did something cool — "AI teaching AI" — becoming the first to have an agent teach Professor Zhiyuan Liu's Tsinghua course "Towards General Artificial Intelligence," to great acclaim among faculty and students.
To verify the scientific validity of AI education, Tsinghua's Institute of Education ran a "real vs. fake Liu Zhiyuan" showdown: a randomized controlled experiment among "the real Liu Zhiyuan," "the AI Liu Zhiyuan in MAIC," and "Liu Zhiyuan in MOOC videos." The result: the AI Liu Zhiyuan won significantly.
Since then, MAIC has been racing ahead:
March 2025: launched on the National Smart Education Platform, and developed "Anything to MAIC," which can draw PPTs, and MaicPBL for project-based learning.
2025 freshman enrollment: MAIC was sent to every new student along with the admission letter. Tsinghua students' first Tsinghua class is an AI class.
Second half of 2025: partnered with more than a dozen secondary schools nationwide to create a new multi-agent classroom where no human teacher lectures and students learn with AI at their own pace.
......
As AI technology matures, its main contribution to education is this: the romantic visions people once had about education can all become reality.
MAIC's product features have also kept iterating with technology:
In early 2024, large language models took off. MAIC applied LLMs and RAG to course production, striving to reduce hallucinations in AI lectures.
In 2025, agent technology advanced and AI coding ability improved greatly. MAIC combined more complex multimodal capabilities to let AI automatically make courseware and build classrooms.
The OpenMAIC open-sourced in 2026 catches the wave of longer-horizon autonomous agents, able to execute more complex and in-depth tasks — for example, the MAIC UI autonomously creates personalized demos to explain specific knowledge according to different time points, student conditions, and learning scenarios, with the agent executing end to end without human intervention.
In the past, many excellent educational ideas were proposed early but could not be implemented due to technology and cost constraints. In the AI era, large models and agents not only provide good tools but liberate traditional education models. The reform and rise of AI education can be seen in Tsinghua's MAIC and OpenMAIC.
Not Just a Tool, but a Paradigm
Over the past two years, the MAIC team has served more than 100,000 teachers in China, gaining a comprehensive understanding of education at different levels and a deeper grasp of its pain points. This cognitive barrier, together with Tsinghua's advanced AI R&D capability, forms OpenMAIC's moat.
One sentence, and your personal tutoring course is generated
OpenMAIC's built-in "Plan agent" proactively plans the syllabus and lesson plans.
For example, if you are a liberal arts student, tell it: "Help me understand Python from scratch; I want to write my first program in 30 minutes." The mini-class OpenMAIC generates is logically coherent and fully in line with a beginner's cognitive patterns:
For example, if you are an AI researcher who needs to quickly learn frontier technology, tell it: "Please explain DeepSeek's latest paper." OpenMAIC automatically structures it around background, bottlenecks, in-class quiz, highlights, hands-on experience, and details, with rich learning content in each part:
One click, and you have a cinematic interactive classroom
OpenMAIC turns dry knowledge points into interactive components you can see, click, and play with.
Slides: supports images, text, and video; export and editing are coming.
Quiz: objective and subjective questions, automatically graded by AI.
Interactive web pages: generate visualized web pages for a scenario — GenUI is awesome.
Want to teach young children about rocks? Enter "My kid loves collecting all kinds of rocks; help me teach some related knowledge in a fun way." It generates an engaging course combining observation, games, and feedback:
Teaching quadratic equations? Flexible curve changes and freeform formula annotation — even high interaction difficulty is no problem:
A group of AI top students to brainstorm with you
Scheduled through LangGraph, OpenMAIC supports multiple agents discussing at a roundtable; you can observe or join at any time.
In a lightweight multi-agent project-based learning interactive web page, students can advance a project through dialogue with agents and complete tasks. Set up a multi-agent discussion about code, and OpenMAIC produces this classroom process:
AI teachers can not only improvise demonstrations on the whiteboard but also hold multi-role discussions around the whiteboard content.
Want to discuss the "Turing test"? AI characters automatically generate diagrams and annotate text on the whiteboard, turning abstract speculation into a visible, participatory multi-agent debate:
Infinite Imagination: From the Classroom to the World
OpenMAIC's boundaries go far beyond the classroom.
Want to analyze the stock prices of Zhipu and MiniMax? It can:
Want to travel into the Three-Body world as the Swordholder and understand the Dark Forest principle in 30 minutes? It can do that too:
This means any scenario that calls for systematic, efficient, interactive learning is a stage for OpenMAIC.
As agent technology keeps breaking through, AI in education is moving from a topic to case after case in practice. And on this land richest in both technology and scenarios, perhaps only China has the chance to give the world an answer for a new education paradigm.
To create a new paradigm, you need to embrace AI deeply enough and have enough experimental scenarios. Globally, only China satisfies both.
"Dare to realize the most beautiful education innovations with the most cutting-edge technology."
➤ Reference links
🔗 https://jcst.ict.ac.cn/article/doi/10.1007/s11390-025-6000-0