Recursive Super Intelligence

Recursive (incorporated as Recursive Superintelligence, Inc., also referred to as Recursive Super Intelligence or RSI) is an artificial intelligence research lab developing recursively self-improving systems intended to automate knowledge discovery. Its stated thesis is that the fastest path to superintelligence runs through AI that improves itself using open-ended algorithms modeled on Darwinian and cultural evolution, in which an archive of diverse discoveries accumulates and each innovation builds on prior ones. The company's first application domain is the science of AI itself — building AI that improves AI — on the reasoning that AI is code and therefore the most immediately tractable target, and that a system able to solve hard AI problems (data efficiency, credit assignment, algorithm and system design) would compound into other scientific disciplines. The company states it prioritizes safety alongside capability.

According to press coverage, the systems under development are meant to automate parts of the research pipeline including model architecture, training methods, evaluation and research direction, using software that generates, tests and refines new capabilities in a continuous cycle with reduced human supervision. The company positions this as an alternative to relying primarily on larger models and more compute. As of May 2026 it was described as a stealth-mode lab preparing to run a first “Level 1” autonomous training system, with a public launch planned for mid-2026; a post titled “First Steps Toward Automated AI Research” appeared on its site in June 2026.

Founding story

Press coverage dates the company's founding to 2025 and describes a founding team drawn from former leaders and researchers at OpenAI, Google DeepMind, Meta AI, Salesforce AI and Uber AI. Named founders include Richard Socher (CEO), Josh Tobin (CTO), Jeff Clune, Tim Rocktäschel, Yuandong Tian, Tim Shi and Alexei Dosovitskiy; investor commentary describes eight founders. Backgrounds cited include Socher's GloVe word-embedding work and his startup MetaMind (acquired by Salesforce, where he served as Chief Scientist); Tobin's time as an early OpenAI employee, founder/CEO of Gantry, and a second OpenAI stint leading Deep Research; Clune's evolutionary and open-endedness research including POET, Darwin Gödel Machines and HyperAgents; Rocktäschel's work on RAG, Rainbow Teaming, Genie, PLR and ACCEL; Tian's decade at Meta including StreamingLM, GaLORE, Coconut and OpenGo; Shi's early OpenAI computer-control research and co-founding Cresta as CTO; and Dosovitskiy's invention of the Vision Transformer. Researchers from Clune's group (Jenny Zhang, Shenghran Hu, Cong Lu) and Rocktäschel's UCL lab (Dominik Schmidt) joined the founding team.

Traction

The company reported a team of more than 25 researchers and engineers as of May 2026, and said it would use new funding to expand compute infrastructure and research operations. It was preparing to run a first "Level 1" autonomous training system, with a public launch planned for mid-2026, and published a post titled "First Steps Toward Automated AI Research" in June 2026.

Latest developments

In May 2026 the company disclosed a $650 million raise at a $4.65 billion valuation led by GV and Greycroft, with AMD Ventures, NVIDIA and Amplify Partners among the investors. On 11 June 2026 it published "First Steps Toward Automated AI Research." It continues to recruit research and engineering staff for its San Francisco and London offices.