Yuandong Tian, who spent more than a decade as a research scientist director inside Meta's Fundamental AI Research lab, has surfaced as a co-founder of Recursive Superintelligence, a startup that emerged from stealth on May 13 with $650 million in funding and a $4.65 billion valuation. The company has fewer than 30 employees and has not shipped a product, yet it has drawn capital from Alphabet's venture arm and two of the largest chipmakers on a single thesis: that artificial intelligence can be engineered to improve itself, faster, in an accelerating loop.
For a reader tracking where frontier AI money is flowing, the bet matters because it reorganizes a whole company around an idea the biggest labs treat only as an internal tool. If recursive self-improvement works, the firm that gets there first could compound its lead rather than grow it step by step — and the people who built today's models are now wagering that the next ones can build themselves.
GV and Greycroft Led the Round, Nvidia and AMD Joined
The financing was led by GV and Greycroft, with Nvidia and AMD participating. The presence of the two companies whose silicon underpins virtually all large-model training is the tell. Strategic checks from the firms that sell the picks and shovels signal that they read self-improving AI as a near-term buyer of compute, not a far-off abstraction. The round was described as heavily oversubscribed, and the company now operates from offices in San Francisco and London.
Recursive is led by Richard Socher, the former Salesforce chief scientist and founder of the search engine You.com, alongside seven co-founders. They include Tian; Tim Rocktäschel, a University College London professor and former DeepMind principal scientist who led world-model and open-endedness research; Alexey Dosovitskiy, a lead author of the Vision Transformer paper that reshaped computer vision; Josh Tobin, formerly of OpenAI; Caiming Xiong, a former Salesforce senior vice president; Tim Shi; and Jeff Clune, known for foundational "AI scientist" research. Peter Norvig, co-author of the standard textbook Artificial Intelligence: A Modern Approach, serves as an adviser.
Tian Left Meta Arguing Small Teams Now Outrun Big Ones
Tian frames his exit after nearly 11 years as a structural bet, not a grievance. Large organizations, he argues, now carry alignment and communication costs that frontier development can no longer absorb: executives push directives down, sprawling teams spend more time syncing with leadership than shipping, and the mismatch shows up as the constant reorganizations and layoffs moving through the industry. A small, low-friction team, in his view, has become the only configuration that keeps pace.
Investors appeared to share that read. Tian has said that in a field changing this fast, top capital is effectively backing people over products, and that Recursive's roster of complementary, accomplished researchers is what justified funding a company with nothing yet to sell. His own record anchors the group: a Shanghai Jiao Tong University graduate with a robotics PhD from Carnegie Mellon, he led the DarkForest Go project and later ELF OpenGo at Meta, then concentrated on reinforcement learning, language-model reasoning, and AI-guided optimization.
👉 Read more:
Hexo Labs Launches an Open-Source Self Improving Agent (SIA)
How Recursive Self-Improvement Works: Models That Rewrite Their Own Training
The mechanism underneath the headline is a closed optimization loop. Instead of human researchers hand-designing each new generation — choosing architectures, gathering data, labeling examples — the system automates parts of its own research and development, and each gain makes it better at producing the next. Tian argues that human stamina has become the binding constraint on model progress, and that handing the discovery process to the system is how a lab raises its rate of discovery rather than its headcount.
Recursive has published a staged path rather than a single leap. The first target is a system with the capability of "50,000 doctors" aimed squarely at automating AI research itself, followed by a "Level 1" autonomous training system with a public launch targeted for mid-2026. To make that loop trustworthy, Tian places unusual weight on interpretability: knowing what a model is doing internally is both the safety floor that catches reasoning drift before it compounds and the efficiency lever that spares a team from burning thousands of GPUs on blind trial and error. If human-level capability is "1" and the goal is "10," he puts current systems at roughly 0.5 to 1 — very early in the climb.
Why Tian Thinks Pretraining, Not Reinforcement Learning, Sets the Ceiling
Several of Tian's technical positions cut against the industry's loudest assumptions, and they explain the company's design choices. He led closely watched work on latent reasoning and contends that chain-of-thought prompting, useful as it is, may not be the most efficient form of machine thinking: a representation in latent space can carry far more information per step, and explore multiple lines at once, than text written out one token at a time. He also argues that reinforcement learning's ceiling is fixed by pretraining — RL mostly promotes a correct answer the model can already produce, so a model that never learned a good internal representation has no winning "seed" for RL to surface. And he holds that language carries a higher ceiling than today's vision and "world model" systems, because language can recombine, self-reference, and redefine concepts in ways pixel statistics cannot.
That skepticism extends to brute-force scaling. Tian's earlier research into "grokking" — the abrupt switch from memorization to genuine generalization once data crosses a critical threshold — informs his warning that the field's exponential appetite for compute and power will eventually meet physical limits. The escape, he argues, is theoretical and efficiency breakthroughs, which is precisely the bet a small lab can make and a hyperscaler chasing the same benchmark cannot.
What Does Recursive Superintelligence Compete Against?
Recursive is not alone in turning models inward, but it is alone in making the loop the product. Anthropic has said most of its code is now written by Claude; OpenAI reported that GPT-5.5 devised a parallelization method that lifted token-generation speed by more than 20%; and Google DeepMind built the AlphaEvolve coding agent for algorithmic discovery. The difference, Recursive's backers argue, is organizational: the major labs sell models and API access and use self-improvement to assist that work, while Recursive is built so the self-improvement engine itself is what it ships.
Tian situates the company inside a wider exodus of senior researchers founding their own labs, and argues the timing is the opening. The late-2025 jump in AI coding tools means a handful of elite engineers can adopt a new production paradigm almost overnight, while a large organization may need a year or more to restructure around it — a window in which newcomers can pull ahead before incumbents adapt.
Tian Says Mid-Level Engineering Work Is Already Replaceable
Tian does not soften the labor implications. He argues that AI can already absorb much of the work done by mid-level engineers, which helps explain why profitable giants keep cutting staff even as revenue holds. He likens the job market to a fish hopping between shrinking puddles: moving from one large employer to another is no escape when the water — the supply of conventional roles — is draining everywhere at once. The durable move, he suggests, is a dimensional jump, toward self-directed work that only a given person would pursue and that generates new data and experience no model has seen. Having predicted after GPT-4's debut that nearly everyone would eventually become an entrepreneur, he casts the next several years as a painful but renaissance-like transition.
Whether the loop delivers runaway acceleration or settles into diminishing returns is still genuinely open. Anthropic co-founder Jack Clark has estimated roughly a 60% chance by the end of 2028, and 30% by 2027, that a system could train a more capable successor without human help. For now the market has priced the possibility at $4.65 billion — for a company four months old that has yet to release a thing.
Frequently Asked Questions
What is Recursive Superintelligence?
Recursive Superintelligence is a San Francisco and London AI startup that emerged from stealth in mid-May 2026 with $650 million in funding at a $4.65 billion valuation. It is building AI systems that can automate their own research and improve themselves in a compounding loop.
Who founded Recursive Superintelligence?
It is led by former Salesforce chief scientist Richard Socher with seven co-founders, including ex-Meta FAIR director Yuandong Tian, former DeepMind scientist Tim Rocktäschel, Vision Transformer author Alexey Dosovitskiy, and former Salesforce executive Caiming Xiong. Peter Norvig is an adviser.
What is recursive self-improvement in AI?
It refers to AI systems that automate parts of their own development, so each improvement makes the system better at producing the next one. The goal is development that compounds rather than depending on human researchers for every new generation.
Why did Yuandong Tian leave Meta?
Tian has argued that large organizations now carry alignment and communication costs that frontier AI work cannot absorb, and that small, low-friction teams move faster. He left after nearly 11 years to co-found Recursive on that premise.
ⓒ 2026 TECHTIMES.com All rights reserved. Do not reproduce without permission.
