Recursive Superintelligence: The $4.65 Billion Bet on AI That Builds I – Good Machine

TL;DR

Who: Eight co-founders from Meta FAIR, DeepMind, OpenAI, and Salesforce — including Richard Socher (CEO), Yuandong Tian, Tim Rocktäschel, Alexey Dosovitskiy (Vision Transformer), Jeff Clune, and Peter Norvig (adviser).
What: Building AI that can recursively self-improve — autonomously handling the full cycle of ideation, implementation, and validation — targeting a system with the capability of "50,000 PhDs" focused on AI research itself.
When: Emerged from stealth on 13 May 2026; founded ~January 2026.
How much: $650M raised at a $4.65B valuation, led by GV and Greycroft, with strategic investment from Nvidia and AMD.
Why it matters: If recursive self-improvement works, it could compress decades of AI progress into months. If it doesn't, it's the most expensive bet ever made on diminishing returns.
Products: Promised in "quarters, not years."

The Emergence

The idea has haunted computer science since the 1960s: an AI system that improves itself, then uses those improvements to improve itself again, faster, in an accelerating loop that eventually outpaces every human researcher on Earth. For most of that time, it remained comfortably theoretical. On 13 May 2026, someone raised $650 million to build it.

Recursive Superintelligence emerged from stealth with a $4.65 billion valuation, a founding team drawn from the uppermost echelons of Meta AI, Google DeepMind, OpenAI, and Salesforce AI, and a thesis that would have sounded like science fiction two years ago but now sits squarely within the Overton window of Silicon Valley ambition. The company is four months old, has fewer than 30 employees, and has not released a product. In the current AI investment climate, the promise of a machine that can improve itself is apparently worth more than many companies that have already built one.

The round was led by GV (Alphabet's venture capital arm) and Greycroft, with "major participation" from Nvidia and AMD — the two chipmakers whose hardware underpins virtually all frontier AI training. Their involvement is telling: strategic investment from the firms that sell the picks and shovels suggests they see recursive self-improvement not as a theoretical curiosity but as a near-term compute customer of extraordinary scale.

The Founding Team: A Super-Lab in Startup Form

The team Recursive has assembled is less a conventional startup and more a research super-lab organised around a single paradigm. Eight co-founders, each with a claim to having shaped the current generation of AI, have converged on the thesis that the next generation can build itself.

Richard Socher — CEO & Co-Founder

Socher is the most commercially recognisable name on the roster. Former Chief Scientist at Salesforce (2017–2020), he founded You.com, the AI-powered search engine that reached a $1.5 billion valuation in 2025. Before that, his doctoral work at Stanford produced some of the most cited papers in natural language processing, and his ImageNet contributions helped define the deep learning era. Socher brings the operational and product instincts of someone who has built and scaled AI companies — and he is emphatic that Recursive is not a "neolab." "I want us to become a really viable company," he told TechCrunch, "to really have amazing products that people love to use."

Yuandong Tian — Co-Founder

Tian's departure from Meta is itself a signal. A graduate of Shanghai Jiao Tong University with a PhD in robotics from Carnegie Mellon, Tian spent over a decade at Meta's Fundamental AI Research (FAIR) lab as a Research Scientist Director. He led the DarkForest Go project — a CNN-based Go AI developed before DeepMind's AlphaGo captured global attention — and later became lead scientist on ELF OpenGo, which open-sourced a reproduction of AlphaZero that defeated professional players 20–0. His work spanned reinforcement learning, LLM reasoning, and AI-guided optimisation. That someone of Tian's pedigree would leave Meta to join a four-month-old startup pursuing the most ambitious goal in the field underscores a broader trend: the talent that built the current generation of AI systems is now betting that the next generation can build itself.

Tim Rocktäschel — Co-Founder

Rocktäschel is arguably the intellectual architect of Recursive's technical approach. A professor of AI at University College London and former Director and Principal Scientist of Google DeepMind's Open-Endedness group, he led the team behind Genie 3, an interactive world model that won the ICML 2024 Best Paper Award. He also pioneered "rainbow teaming" — a technique where two AIs co-evolve in adversarial loops, one attacking and the other defending, generating millions of safety-testing scenarios that no human team could produce. The technique is now used across all major AI labs. Rocktäschel operates from Recursive's London office, which also serves as the company's legal home, incorporated at the end of 2025.

Alexey Dosovitskiy — Co-Founder

Dosovitskiy is a co-author of "An Image is Worth 16x16 Words" (2020), the paper that introduced the Vision Transformer (ViT) and fundamentally reshaped computer vision research. ViT demonstrated that transformer architectures — originally designed for natural language — could outperform convolutional neural networks on image tasks, a paradigm shift now embedded in virtually every frontier multimodal model. His presence on the team signals that Recursive's ambitions extend beyond language into the full spectrum of AI capabilities.

Jeff Clune — Co-Founder

Clune is one of the world's foremost researchers on open-endedness and AI-generating algorithms. Formerly head of OpenAI's Open-Endedness team, a professor of computer science at the University of British Columbia, and senior research manager at Uber AI Labs, Clune's academic work — including the seminal Darwin Gödel Machine paper — provides much of the intellectual foundation for Recursive's approach. Four of Recursive's employees co-authored that paper alongside him, reflecting a deep, pre-existing intellectual lineage within the team.

Josh Tobin — Co-Founder

Tobin was one of the first researchers at OpenAI, where he worked on robotics and large-scale reinforcement learning before leading the Codex and deep research teams. He later founded Gantry, a data-quality platform for machine learning systems. Tobin brings the rare combination of frontier research credentials and hard-won operational experience in shipping AI products.

Tim Shi — Co-Founder

Shi was a research scientist at OpenAI working on agentic and code-generation systems, and later co-founded Cresta, the AI-powered contact centre platform that reached unicorn status. His experience bridging research and commercial deployment reinforces Recursive's insistence that it is building a product company, not merely a research lab.

Caiming Xiong — Co-Founder

Xiong was previously at Salesforce AI, where he led research in natural language processing and enterprise AI systems. His expertise in deploying AI at enterprise scale complements the team's deep research bench.

Advisory Board

Peter Norvig, co-author of Artificial Intelligence: A Modern Approach — the standard university textbook in the field — and former Director of Research at Google, serves as an adviser. His involvement lends institutional gravitas to an already formidable team.

The Thesis: Recursive Self-Improvement as a Product

The core thesis behind Recursive is deceptively simple, and GV's Tom Hulme captured it succinctly: "AI is code, and now AI can code. When these two realities connect, the self-improvement loop can be closed."

But Socher is careful to distinguish recursive self-improvement from mere automated research. "A lot of people already assume it happens when you just do auto-research," he told TechCrunch. "You can take AI and ask it to make some other thing better, which could be a machine learning system, or just a letter that you write. But that's not recursive self-improvement. That's just improvement."

True recursive self-improvement, in Recursive's formulation, means the entire process of ideation, implementation, and validation of research ideas becomes automatic. The system identifies its own weaknesses, writes its own benchmarks, designs experiments to address shortcomings, executes those experiments, validates the results, and integrates the improvements — all without human intervention. Then it repeats the cycle, using its newly enhanced capabilities to generate still better improvements.

The Technical Approach: Open-Endedness

The company's technical differentiator is its embrace of "open-endedness" — a concept drawn from evolutionary biology and pioneered in AI by Rocktäschel and Clune. In biological evolution, organisms adapt to their environment, and others counter-adapt to those adaptations, producing an endless cascade of innovation without any predefined endpoint. Eyes, wings, and immune systems emerged not because they were specified in advance, but because the process of variation and selection, iterated over millions of generations, discovered them.

Recursive aims to instantiate the same principle in code. Rather than training AI systems against fixed objectives set by human researchers, the company is building architectures where AI systems co-evolve through iterative self-challenge. Rocktäschel's rainbow teaming provides a template: two AIs locked in adversarial loops, one attacking and the other defending, generating millions of scenarios that drive both systems to higher levels of capability.

Socher describes the ambition in evolutionary terms: "In biological evolution, animals adapt to the environment, and then others counter-adapt to those adaptations. It's just a process that can evolve for billions of years, and interesting stuff keeps happening. That's how we developed eyes in our heads."

The Roadmap

Recursive has outlined a staged roadmap. The first milestone is to train a system with the capabilities of "50,000 PhDs" — a phrase that appears in both the company's materials and GV's investment thesis — focused initially on automating AI scientific research itself. From there, the company plans to run what it calls a "Level 1" autonomous training system, with a public launch targeted for mid-2026.

The funding will be used in significant part to secure the large-scale compute infrastructure required to run these experiments. Socher has indicated that products will ship in "quarters, not years," and suggested that internal progress has been faster than initially projected.

Once the self-improving engine is running on AI research, the company intends to point its "Eureka machine" at broader scientific domains. "We will start with AI research itself but eventually hope to expand its aperture to physics, chemistry and especially pre-clinical biology," Socher wrote on X. "AI will be to biology what calculus was to physics — a new language and way of thinking that deals with complex systems and helps us understand and engineer them better."

The Competitive Landscape: A Race Already Underway

Recursive Superintelligence is not pursuing this thesis in isolation. The largest AI laboratories are already using their own models to accelerate research, and the boundary between AI-assisted and AI-driven research is blurring rapidly.

Anthropic has disclosed that the majority of its code is now written by Claude. OpenAI reported that GPT-5.5 developed a parallelisation method that boosted token generation speeds by more than 20% — an optimisation discovered by the model itself, not by human engineers. Google DeepMind has built AlphaEvolve, a coding agent designed for scientific and algorithmic discovery. Google co-founder Sergey Brin has reportedly described coding gains as a path to "AI takeoff" internally. OpenAI CEO Sam Altman has said the company hopes to have a system capable of doing the work of a "less experienced" researcher by the northern autumn of 2026.

What distinguishes Recursive from these efforts is structural. None of the major laboratories has organised an entire company around recursive self-improvement as its core commercial thesis. OpenAI, Anthropic, and Google DeepMind all use AI to assist their research workflows, but their businesses are built around selling models and API access. Recursive is betting that the self-improvement loop itself is the product — and that the first company to close that loop will enjoy a compounding advantage that no competitor can match.

The competitive field extends beyond the major labs. Ineffable Intelligence, founded by former DeepMind researcher David Silver, raised a record $1.1 billion seed round in April 2026 to pursue superintelligence through reinforcement learning. Ricursive Intelligence (note the spelling — a different company) spun out of Alphabet to commercialise AI-assisted chip design. The broader "neolab" category — startups that prioritise research over immediate product revenue — has attracted billions in venture funding over the past twelve months.

The Investor Logic

GV's investment thesis, articulated by General Partner Tom Hulme, frames Recursive as a bet on an "orthogonal S-curve." The standard laws of pre-training are delivering diminishing returns at the frontier; to achieve true reasoning and scientific discovery, the industry needs fundamentally new approaches. Recursive's open-ended architecture represents one such approach.

The presence of Nvidia and AMD on the cap table is strategically significant. Both companies sell the GPUs and accelerators on which all frontier AI training depends. Their investment suggests they view recursive self-improvement as a near-term driver of compute demand — and potentially an exponential one. If Recursive's thesis is correct, a self-improving AI system would consume ever-increasing quantities of compute as it runs faster and more ambitious experiments. The chipmakers are positioning themselves at the source of that demand.

The round was described as heavily oversubscribed. The $650 million final figure — up from the $500 million initially reported by the Financial Times in April — reflects the intensity of investor appetite for frontier AI exposure, particularly when the founding team carries the credentials Recursive has assembled.

The Open Questions

Whether Recursive's bet pays off depends on a question that remains genuinely open: does recursive self-improvement produce the kind of runaway acceleration its proponents describe, or does it converge on diminishing returns as each cycle of improvement yields smaller gains?

Anthropic co-founder Jack Clark has estimated a roughly 60% probability that a system capable of training a more powerful successor on its own, without human involvement, will exist by the end of 2028, and a 30% chance by 2027. These are not long-range forecasts; they are near-term operational assumptions.

Sceptics point out that current AI systems remain far from the point where humans can be removed from the loop. Strictly defined recursive self-improvement requires a system to improve itself without human instruction — setting its own goals, defining its own success criteria, and deciding which changes to retain. Today's systems, for all their capabilities, still rely on humans to set objectives and evaluate outcomes.

The Bigger Picture

Recursive Superintelligence's launch is more than a single company's funding announcement. It is a milestone in a structural shift in how the AI industry organises itself.

Over the past eighteen months, the frontier of AI research has increasingly migrated from large corporate labs to venture-backed startups. Former Meta, Google, and OpenAI researchers have founded a wave of new companies — Ineffable Intelligence, Thinking Machines Lab, AMI Labs, and now Recursive — each pursuing a different path toward advanced AI. The talent that built the current generation of systems is betting, with billions of dollars of venture capital behind them, that the next generation requires a different organisational form.

Recursive's thesis — that the self-improvement loop itself is the product — represents the most radical version of this bet. If it works, the implications extend far beyond AI research. A system that can automate the scientific method could, in principle, be directed at any quantitative problem: drug discovery, materials science, climate modelling, fusion physics. Socher's "Eureka machine" is, at its limit, a general-purpose engine for scientific discovery.

For now, what is certain is the price the market has placed on the possibility. Four months old. Fewer than 30 employees. No product. $4.65 billion. In May 2026, the promise of a machine that can improve itself is one of the most valuable ideas in the world.

Sources

The New York Times — "A. I. Start-Up Hopes to Build a Self-Improving Machine" (13 May 2026). Cade Metz. nytimes.com
GV (Google Ventures) — "Why We Invested in Recursive Superintelligence" (13 May 2026). Erik Nordlander. gv.com
South China Morning Post — "AI start-up Recursive Superintelligence raises US$650 million at US$4.65 billion valuation" (14 May 2026). Ben Jiang. scmp.com
The Information — "GV, Greycroft Lead $650 Million Round for AI Startup Recursive Superintelligence" (13 May 2026). Stephanie Palazzolo. theinformation.com
Recursive Superintelligence — Official website. recursivesuperintelligence.ai
Jeff Clune et al. — "AI-GAs: AI-generating algorithms, an alternate paradigm for producing general artificial intelligence" (2020). Nature Machine Intelligence, 2, 44–52. nature.com
Jeff Clune — "A Darwin Gödel Machine: Towards Self-Replicating, Self-Improving AI" (2023). arxiv.org
Ineffable Intelligence — "Introducing Ineffable Intelligence" (April 2026). ineffable.ai
Anthropic — "Core Views on AI Safety" (2025). anthropic.com
OpenAI — "Planning for AGI and Beyond" (2023). openai.com