Recursive Co-founder Tian Yuandong; image processed by AI
By Li Hailun
Edited by Xu Qingyang
After leaving Meta, Tian Yuandong has taken on a new role.
In October 2025, Meta laid off about 600 people from its AI research division, including Tian Yuandong, then Research Scientist Director at FAIR (the Fundamental AI Research team). When the news broke, AI companies like OpenAI and Anthropic scrambled to court him. But Tian didn't join any of them; instead, he chose to start his own venture.
On May 13, US local time, he emerged as a co-founder of Recursive Superintelligence, announcing the move on X. The company, founded by several key figures who previously held roles at major AI firms, officially launched that day.
A post from Recursive co-founder Tian Yuandong
Just six months old with a team of fewer than 30, the company has already secured $650 million in funding at a staggering $4.65 billion valuation. The round was led by GV (Google Ventures) and Greycroft, with participation from AMD Ventures and Nvidia. The funds will be used to scale computing infrastructure and operations in San Francisco and London.
Tian stated on X that the team is building an AI capable of autonomously discovering knowledge and recursively self-improving, which will fundamentally change how science and technology advance.
01
Building an AI "Dream Team"
Behind Recursive is a truly stellar list of co-founders.
The company was established by eight co-founders, including Tian Yuandong, former Salesforce Chief Scientist Richard Socher, Caiming Xiong, Tim Shi, and Josh Tobin. They previously held research leadership or core positions at top labs like OpenAI, Google DeepMind, Meta, and Salesforce.
The team's expertise spans key areas such as agentic AI scientists, architecture and algorithm design, world models, optimization, and interpretability.
Many of them are pioneers in their respective fields, having led or significantly contributed to major advances like open-ended algorithms, quality diversity algorithms, generative AI algorithms, self-improving coding agents, automated red-teaming and capability discovery, foundational world models, deep learning, Vision Transformers, and Retrieval-Augmented Generation.
Recursive's eight co-founders: Alexey Dosovitskiy, Caiming Xiong, Jeff Clune, Josh Tobin, Richard Socher, Tim Rocktaschel, Tim Shi, and Tian Yuandong
Now, this group has come together for a shared belief: achieving recursive self-improvement in AI.
02
Writing "Evolution" into Code
The theory behind recursive self-improvement isn't complicated. Recursive Co-founder and CEO Richard Socher explained, "AI is code. And now, AI can code. All the pieces are in place."
The logic is that since AI itself is composed of code, and current AI possesses the ability to write code, it makes perfect sense to have AI improve its own code.
In his view, the path of machine learning over the years is clear: more compute and more data mean that methods once painstakingly designed by humans are being progressively replaced by AI-driven methods. Recursive aims to follow this path to its logical conclusion.
Recursive Co-founder and CEO Richard Socher
To explain the company's technical philosophy, Recursive's announcement begins with the birth of natural intelligence. Human intelligence was co-created by two open-ended processes: Darwinian biological evolution and cultural evolution. These processes continuously accumulate archives of interesting and distinct discoveries, with every new invention building upon prior achievements.
Biological evolution shaped our bodies, vision, and simple reflexes, while cultural evolution developed reasoning, language, and science atop that hardware. This process has no ceiling; innovation continues indefinitely.
Recursive believes that AI science follows this same pattern of open-ended innovation, except that, until now, discoveries were driven by human scientists. Now, it's time to pass the baton to AI itself.
Therefore, Recursive asserts that the fastest path to superintelligence will be achieved by AI capable of recursive self-improvement, and this improvement must rely on open-ended algorithms that can drive endless innovation. They plan to start with AI's own scientific research, creating an AI that can improve AI, and then rapidly expand this methodology to a broader range of scientific fields.
Socher wrote on social media: "AI is to biology what calculus was to physics — a new language and way of thinking for handling complex systems, helping us better understand and design them."
Recursive's official announcement, stepping out of "stealth mode" to launch publicly
03
An AI That Researches and Improves Itself
So, how exactly would such a self-improving AI system work?
According to information disclosed by the company, the AI developed by Recursive will simulate "an open-ended, automated scientific discovery process," proposing its own experimental ideas, running tests, and verifying results.
Its scope for improvement goes far beyond general code optimization. It includes not only improving its own code but also upgrading its "tool suite" — the auxiliary programs AI providers use to enhance algorithmic outputs.
Furthermore, the system will also seek ways to improve its training and inference infrastructure. In Recursive's plan, these self-improvement experiments can continuously generate and optimize new capabilities in an open-ended loop, independently discovering better learning methods without requiring constant human oversight.
Of course, safety is paramount throughout this process. Recursive has made it clear that it will set up safety guardrails to ensure the system helps humanity flourish by maximizing benefits and minimizing risks, preventing the software from producing risky outputs.
For Recursive, this long journey towards superintelligence has only just taken its first step.