After leaving Meta, Tian Yuandong has just officially announced his entrepreneurship.

After leaving Meta, Tian Yuandong also started his own business.

Just now, the startup Recursive_SI officially made its debut and announced the list of founders, including Tian Yuandong.

In addition to Tian Yuandong, the founding team also includes Richard Socher (CEO), Tim Rocktäschel, Jeff Clune, Tim Shi, Caiming Xiong, Alexey Dosovitskiy and others.

These founding members have participated in establishing the AI research labs of Salesforce and Uber, and have held leadership positions in teams such as OpenAI, DeepMind, Google Brain and Meta, with rich scientific research and entrepreneurial experience.

Recursive_SI is committed to creating an artificial intelligence that can conduct experiments autonomously and improve itself in a safe way - evolving continuously in an open-ended automated scientific discovery process, which is considered the most likely path to superintelligence.

Currently, Recursive has raised $650 million and is valued at $4.65 billion. The investment is led by GV (Google Ventures) and Greycroft, and AMD Ventures and NVIDIA have also made significant investments.

The team has more than 25 members and is still expanding. The team has attracted many outstanding talents, including Zhuge Mingchen who is about to join.

Zhuge Mingchen is now a founding member of Recursive. He graduated with a doctorate in computer science from King Abdullah University of Science and Technology (KAUST), under the guidance of Professor Jürgen Schmidhuber, known as the "father of LSTM". His research focuses on coding agents, recursive self-improvement (RSI) and next-generation machine paradigms.

Since 2023, Zhuge Mingchen has been systematically exploring the direction of Recursive Self-Improvement (RSI).

During the MetaGPT period, he proposed that agents should have a mechanism for continuous self-optimization and ability evolution, and has been continuously promoting this research route in subsequent work. Among them, GPTSwarm is considered one of the earliest RSI system paradigms in the LLM era. It was the first to systematically propose and verify a self-organizing collaborative framework based on graph-based agents, realizing collaboration, feedback and ability evolution among agents through a dynamic graph structure. Its core idea has been widely adopted by many subsequent multi-agent and Agentic AI works. Agent-as-a-Judge further explores the continuous feedback and self-evaluation mechanism in long-term tasks, trying to solve the problems of continuity and stable optimization of agents in complex tasks. The research on NeuralComputer further aims at the next-generation AI system architecture, exploring a new machine paradigm that integrates memory, reasoning and autonomous evolution capabilities.

It can be seen that the research team joining Recursive has profound academic experience in the direction of recursive self-improvement.

Many founders, including Tian Yuandong, have promoted it on X: We are creating an artificial intelligence that can automatically discover knowledge and recursively improve itself - this open-ended process will fundamentally change the way science and technology progress.

The team is at the forefront of the industry in multiple core areas of recursive self-improvement artificial intelligence.

The members have made significant breakthroughs in areas such as open-ended algorithms, quality diversity algorithms, AI generation algorithms, self-improving programming agents, automated red team testing and ability discovery, prompt engineering and its automation, learning challenges and environment generation, basic world models, deep learning in natural language processing, visual Transformers, retrieval-augmented generation, and AI scientists.

Therefore, we are really looking forward to the next research of Recursive_SI.

This article is from the WeChat official account "Machine Intelligence". The author is Machine Intelligence, and the editor is the editorial department of Machine Intelligence. It is published by 36Kr with authorization.