On October 22, 2025, Meta CEO Mark Zuckerberg approved a layoff order.

600 employees from the AI department were laid off, and even the core team of FAIR (Facebook AI Research) was not spared. Tian Yundong and his team were all swept up.

Ironically, just nine months earlier, Meta was rushing to have them fight fires—less than two months before the release of Llama 4, Tian Yundong’s team was forcibly pulled from fundamental research to support the generative AI product line’s post-training and bug fixes.

The people who should really be solving the problems were laid off, while those truly responsible were not on that list.

In January 2026, Tian Yundong and seven other top AI researchers co-founded Recursive Superintelligence.

In May 2026, the company officially announced: completed $650 million in funding, with a valuation of $4.65 billion. GV and Greycroft led the investment, with AMD Ventures and Nvidia participating.

Meta probably didn’t expect that what they personally sent away was not just one person.

On the day of layoffs, the entire AI community was fighting over him.

As soon as Tian Yundong’s resignation tweet went out, the comment section instantly turned into a high-end Silicon Valley “Boss Zhipin” scene.

OpenAI, xAI, Anthropic, ByteDance, Google DeepMind… all the big AI companies you can name, they’re all here.

One opportunity after another, wealth and glory right in front of him.

Tian Yundong waved his hand and refused all of them.

Many people don’t understand: such a great opportunity, why would he give it up?

But if you know his background, you’ll understand—people like him are born not to be employees.

Born in Shanghai, undergraduate and master’s degrees from Shanghai Jiao Tong University, PhD in Robotics from Carnegie Mellon University (CMU). Joined Meta FAIR in 2013, nearly ten years of service. Reinforcement learning, multi-agent systems, large model reasoning and efficiency optimization, deep learning theoretical analysis—all cutting-edge, hard-core AI directions.

He is the kind of researcher “focused on theory, bottom-layer, and difficult problems.” He may not deliver flashy demos quickly, but he determines the technological ceiling of a company for the next three to five years.

Having him laid off is not his loss, but Meta’s.

Eight co-founders, forming an AI Avengers Alliance

Tian Yundong didn’t go to any big tech company. He chose a tougher path: to be his own boss.

In January 2026, Recursive Superintelligence was registered in the UK. The founding team is dubbed the “Silicon Valley AI Dream Team”—eight co-founders, almost covering the most core research institutions across the AI industry chain.

● Richard Socher (CEO)—former Salesforce Chief Scientist and EVP, founder of search engine You.com. One of the key figures who truly brought neural network methods into NLP, with over 180k citations on Google Scholar.

● Tim Rocktäschel—Head of Open-Ended Intelligence at DeepMind, UCL professor, whose Rainbow Teaming method has become an industry standard in AI safety.

● Shi Tianlin (Tim Shi)—Tsinghua Yao Class alumnus, former OpenAI researcher, co-founder and former CTO of AI unicorn Cresta.

● Alexey Dosovitskiy—First author of the Vision Transformer (ViT) paper, which completely changed the paradigm of computer vision research.

● Caiming Xiong—Former head of Salesforce AI research, led multimodal pretraining research.

● Jeff Clune—Former OpenAI researcher, pioneer in open-endedness and quality diversity algorithms.

● Tian Yundong—Former Director of Research Science at Meta FAIR, expert in reinforcement learning and multi-agent systems.

Additionally, Google’s former research director Peter Norvig, author of the AI textbook Artificial Intelligence: A Modern Approach, also joined as an advisor.

Eight people from OpenAI, Google DeepMind, Meta AI, Salesforce AI, Uber AI. This is not just a startup team; it’s an AI Avengers Alliance.

The company currently has only 25 people, with dual headquarters in San Francisco and London, covering core areas like agent-based AI, algorithm architecture, world models, and interpretability. The classic Silicon Valley narrative: using the fewest people to bet on the furthest future.

Recursive Self-Improvement: AI self-upgrading?

Recursive. The company’s name embodies its technological philosophy.

They are betting on “Recursive Self-Improvement.” The core idea: build an AI system capable of autonomous scientific discovery—hypothesis generation, experiment design, result evaluation, iterative optimization—in an open-ended cycle of self-evolution.

The current large model race still fundamentally follows the Scaling Law logic: bigger models, more data, stronger compute. This path has indeed brought explosive breakthroughs, but marginal returns are diminishing, while training costs are skyrocketing exponentially. The industry is anxiously asking: after the big models, where will the next capability leap come from?

Recursive’s answer: break out of the Scaling Law, let AI itself develop its tech tree.

CEO Socher’s explanation is very clear: “AI is code itself. Now, AI can write code. The necessary elements are all in place.”

Their roadmap involves two steps: first, train a system with the “capability of 50,000 PhDs” to automate AI scientific research—gradually removing human researchers from the cycle of experiment design, paper reading, and hypothesis validation; second, extend this recursive optimization mechanism into broader fundamental sciences like drug discovery, battery materials, and nuclear fusion physics.

It’s not about building a smarter chatbot, but enabling AI to self-evolve. If successful, this will far surpass just releasing another LLM.

Meta’s departure was not just about losing a person

Tian Yundong’s story is more than just a personal comeback narrative.

It reflects the dilemma faced by big Silicon Valley companies—shifting from long-term fundamental research to short-cycle product delivery, from patiently refining technology to rushing to meet quarterly KPIs.

Those who truly determine the technological ceiling for the next three to five years are laid off, while those responsible for strategic mistakes are not on the layoff list.

This “Xerox PARC moment” has played out repeatedly in tech history: Bell Labs disbanded its top fundamental physics team, and the transistor revolution happened elsewhere. Xerox PARC’s graphical interface was “borrowed” by Steve Jobs, leading to Apple and Microsoft becoming trillion-dollar companies. And Xerox itself?

The logic of big companies is quarterly financial reports; the logic of science is planting trees for ten years. When these two conflict, the sacrificed always is the latter—and the market ultimately rewards those willing to tackle hard problems.

The 25 people at Recursive may be writing the beginning of the next decade. Meta is cutting costs, but the market rewards value.

The ultimate form of AI is not a smarter tool, but a species capable of self-upgrading its tech tree.

From being laid off in October 2025 to announcing $650 million in funding in May 2026—Tian Yundong did it in less than seven months.

This is not just a feel-good story; it’s the real script of Silicon Valley.

Sometimes, being “optimized out” is actually the true start of your life.

Big tech companies cut costs; the market rewards value. When everyone chases short-term KPIs, those willing to tackle hard problems and bet on long-term vision become the most scarce.

Recursive’s first “L1-level autonomous training system” plan will debut in mid-2026. By then, Meta may face that familiar question again:

Why can they always send out the truly worth betting on at the most critical moments?