Recursive Superintelligence
Last reviewed
Jun 3, 2026
Sources
11 citations
Review status
Source-backed
Revision
v1 · 1,709 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 3, 2026
Sources
11 citations
Review status
Source-backed
Revision
v1 · 1,709 words
Add missing citations, update stale details, or suggest a clearer explanation.
Recursive Superintelligence is an artificial intelligence research lab that came out of stealth on May 13, 2026 with roughly $650 million in funding at a reported $4.65 billion valuation [1][2][3]. The company was founded in early 2026 by a group of eight researchers led by Richard Socher, the former chief scientist of Salesforce and founder of the search startup You.com, and its stated goal is to build recursively self-improving AI: systems that improve their own design without human help [2][4]. Its slogan, "AI that builds AI," captures a bet that the fastest route to superintelligence runs through automating AI research itself [3][5]. The lab operates out of San Francisco and London and had fewer than 30 employees at launch [1][4].
Despite a near identical sounding name, Recursive Superintelligence is a separate company from Ricursive Intelligence, the chip design AI startup founded by Anna Goldie and Azalia Mirhoseini. The two share a thematic interest in self-improving systems and, by coincidence, similar valuations, but they have different founders, locations, and product focus.
Recursive Superintelligence was incorporated in London and was about four months old when it exited stealth, which places its founding in roughly January 2026 [1][4]. Socher serves as chief executive. Before this venture he was Salesforce's chief scientist, an early author on the ImageNet project, and the founder and CEO of You.com, an AI search company that reached a $1.5 billion valuation [2][6].
The other co-founders are an unusually senior group drawn from Google DeepMind, OpenAI, and Meta AI's Fundamental AI Research lab (FAIR) [1][3]. The list reported across launch coverage and confirmed by lead investor GV is:
| Co-founder | Prior affiliation |
|---|---|
| Richard Socher (CEO) | Chief scientist, Salesforce; founder, You.com [2] |
| Tim Rocktäschel | Professor, University College London; led the Open-Endedness team at Google DeepMind [4][5] |
| Yuandong Tian | Research scientist director, Meta FAIR [2][7] |
| Alexey Dosovitskiy | Co-author of the Vision Transformer (ViT); Google Brain and DeepMind [3][7] |
| Josh Tobin | OpenAI, where he led work on Codex and deep research [3][5] |
| Caiming Xiong | Salesforce AI Research [7] |
| Tim Shi | Co-founder of Cresta; formerly OpenAI [5][7] |
| Jeff Clune | University of British Columbia and DeepMind; known for AI-generating algorithms [5][7] |
Peter Norvig, co-author of the standard textbook Artificial Intelligence: A Modern Approach, is listed as an adviser to the company [3][5].
The roster is notable for how concentrated it is around one research subculture. Rocktäschel and Clune both spent years on open-endedness, the study of algorithms that keep generating novelty indefinitely rather than converging on a single fixed objective. Rocktäschel ran the Open-Endedness team at DeepMind and worked on world models including Genie, while Clune is widely associated with the idea of "AI-generating algorithms" and co-authored work such as The AI Scientist on automating research [5]. Bringing that group together with applied builders like Tobin and Tian, and with Socher's company building experience, is the lab's main asset. It had no product at launch.
The company's central claim is that recursive self-improvement is the shortest path to superintelligence, and that the way to get there is through open-ended algorithms. As the company put it in its launch statement, "The fastest path to superintelligence will be realised by AI that recursively improves itself, and does so via open-ended algorithms that drive endless innovation" [4][8].
Socher has been careful to draw a line between ordinary model improvement and the thing the lab is actually chasing. In an interview he argued that a model getting better at a benchmark is not the goal: "Our unique approach is to use open-endedness to get to recursive self-improvement, which no one has yet achieved," he said, adding that a system that simply improves on a fixed metric is "just improvement," not the recursive kind [5]. The recursive version requires the full research loop, ideation, implementation, testing, and refinement, to run autonomously, so that each cycle produces a system capable of running the next cycle faster [5][9].
GV, which co-led the round, framed the bet in a single line: "AI is code, and now AI can code. When these two realities connect, the self-improvement loop can be closed" [9]. In practice the lab describes a system that can read its own code, find its own weaknesses, write new benchmarks and approaches to address them, test the results, and rewrite parts of its own training, evaluation, and infrastructure [4][9]. One concrete near term target the company has described is training a system with the research capability of roughly "50,000 PhDs" pointed at automated scientific discovery [9].
The open-endedness piece is what distinguishes the approach from a simple optimization loop. Instead of handing the system one objective, the lab wants it to keep generating new environments, problems, and forms of adaptation on its own, an idea borrowed from biological and cultural evolution [4][5]. The same framing extends to safety: the lab points to techniques like rainbow teaming, where one model continuously probes another to surface failure modes, as a way to replace some manual red teaming with an adversarial process that co-evolves with the system it is testing [5].
Socher has said the lab will "start with AI research itself but eventually hope to expand its aperture to physics, chemistry and especially pre-clinical biology" [2]. On timing he has avoided the open ended research lab posture, telling reporters that "there will be products, and you'll have to wait quarters, not years" [5]. Coverage of the launch pointed to a first autonomous training run and a public launch targeted for mid 2026 [1][7].
The $650 million round was described as heavily oversubscribed and valued the company at $4.65 billion [3][8]. It was co-led by GV (Alphabet's venture arm, formerly Google Ventures) and Greycroft, with participation from the venture arms of chipmakers Nvidia and AMD [1][2][3]. Having both Nvidia and AMD on the same cap table is unusual, since the two are direct competitors in AI accelerators, and several outlets singled it out as a sign of how much strategic interest the founding team attracted [10].
The headline figure firmed up over a few weeks. In April 2026 the Financial Times and others reported the company was raising around $500 million at roughly a $4 billion valuation in a round led by GV with Nvidia participating [11]. By the May 13 stealth exit the round had grown to $650 million at $4.65 billion [2][3].
| Detail | Value |
|---|---|
| Announced | May 13, 2026 [1][3] |
| Total raised | ~$650 million [2][3] |
| Valuation | ~$4.65 billion [2][3] |
| Co-lead investors | GV, Greycroft [1][3] |
| Strategic investors | Nvidia, AMD (venture arms) [1][2] |
| Adviser | Peter Norvig [3][5] |
GV partner Tom Hulme, explaining the firm's decision, wrote that the company explores "orthogonal S-curves" needed to complement standard pre-training and reach genuine reasoning and scientific discovery, and added: "There are few companies where the saying 'the possibilities are endless' actually holds true, but this is one of them" [9]. The company has said it will use the capital to hire, to build out training and inference infrastructure, and to fund the compute needed for open-ended automated research [4][9].
The company is headquartered across San Francisco and London, with the legal entity incorporated in the United Kingdom [1][4]. At launch it reported a team of more than 25 researchers and engineers, described in some coverage as "fewer than 30," and no released product [1][4]. The transatlantic structure mirrors the founders' own geography, with Rocktäschel based around UCL in London and much of the rest of the team in the San Francisco Bay Area [1][4].
Recursive Superintelligence arrived during a stretch in which a handful of small, heavily funded labs raised enormous sums on the strength of their founders alone, often before shipping anything. The pattern echoes Safe Superintelligence, the lab Ilya Sutskever started in 2024, and Mira Murati's Thinking Machines Lab, both of which raised billions with no public product [10]. Commentators used Recursive's raise as a clean example of AI research talent itself becoming a venture asset, with the round priced largely on the team's track record in reinforcement learning, open-endedness, and large model research [10].
The thesis also sits inside a broader 2026 conversation about automated AI research and whether AI systems can meaningfully accelerate their own development. Other groups, including the separately named Ricursive Intelligence, pursued related feedback loop ideas in narrower domains such as chip design, while large labs experimented with AI assisted research agents [10]. What set Recursive apart was the explicit framing of recursive self-improvement as the product rather than a side effect, a stance that drew both interest and unease given the long running debate about whether such a loop could accelerate past human oversight. Whether the lab can turn a strong founding team and a clear thesis into a working self-improving system remains the open question, and the company itself acknowledges that no one has built one yet [5].