At a glance: Recursive Superintelligence. Stealth exit: May 13, 2026. Funding: $650 million. Valuation: $4.65 billion. Investors: GV and Greycroft (co-leads), NVIDIA, AMD Ventures, and others. Employees: fewer than 30. Product: none released. Goal: build AI systems that autonomously identify weaknesses, redesign their own architectures, and improve themselves in a closed loop. Offices in San Francisco and London. Part of our AI Models & Companies reviews.
Most AI startups in 2026 are racing each other to build a better chatbot or a faster coding assistant. Recursive Superintelligence is not doing that.
The company, which came out of stealth on May 13 after roughly 18 months of quiet recruiting, is pursuing something that the rest of the industry officially classifies as a long-term research goal: AI that rewrites itself. Systems that don’t just get trained by engineers — systems that identify their own architectural limitations and fix them, autonomously, in an accelerating loop.
Eight co-founders. $650 million. Under 30 employees. And a specific timeline for when they expect the loop to start running.
The Team
The founding group is unusual for being simultaneously overcredentialed and unusually coherent in what they’re trying to do.
Richard Socher (CEO) is the former chief scientist at Salesforce and founder of the AI search engine You.com. He is one of the figures credited with popularizing transfer learning for NLP before the transformer era.
Yuandong Tian was a research director at Meta’s Fundamental AI Research (FAIR) lab, where he led work on reinforcement learning, long-context reasoning in LLMs, and AI-guided optimization. He is one of the co-inventors of several techniques now standard in foundation model training.
Alexey Dosovitskiy is the lead author of the Vision Transformer (ViT) paper — one of the most cited machine learning papers of the last decade. ViT’s approach of applying transformer architectures to image patches is now the default for vision models across Google, OpenAI, and most large labs.
Tim Rocktäschel was a principal scientist at Google DeepMind and a professor of AI at University College London. He directed DeepMind’s Open-Endedness research group, which produced foundational work on AI systems that generate their own curricula.
Josh Tobin is a former OpenAI researcher who did significant work on simulation-to-real transfer and robust AI deployment.
Jeff Clune ran OpenAI’s open-endedness research program before moving to UBC. His lab’s work on systems like POET and AI-GAs — evolutionary algorithms that generate increasingly complex environments and agents in tandem — is directly relevant to what Recursive Superintelligence is building.
Caiming Xiong was a VP of Salesforce Research and the head of its NLP and AI group.
Tim Shi is a co-founder and engineer with prior experience at Cresta and connections to the Salesforce Research orbit.
The common thread across nearly every co-founder is a background in open-ended learning, reinforcement learning, and systems that improve without being explicitly told what to improve next.
The Thesis
Current AI systems — including GPT-5.5, Claude 4, and Gemini 3.5 — improve through a training cycle that requires human-designed objectives, human-curated data, and human-run infrastructure. Recursive Superintelligence’s founding premise is that this model will eventually be superseded by systems that can run their own training cycles: identifying failure modes, generating new training data to address those failure modes, and modifying their own architectures accordingly.
This is sometimes called recursive self-improvement, and it occupies a peculiar position in the AI research landscape — widely agreed to be theoretically possible, widely assumed to eventually happen, and widely avoided as a near-term engineering goal because the safety and alignment implications are severe.
Recursive Superintelligence is not avoiding it. The company is building toward it explicitly, with GV, Greycroft, NVIDIA, and AMD Ventures providing the capital.
The Roadmap
The company has laid out a three-stage plan:
Stage 1 — The “50,000 doctors” system. The first milestone is training a model with the integrated medical and scientific reasoning capacity equivalent to a large panel of domain experts. The framing is deliberately provocative — it is meant to communicate the scale of expertise compression the company believes is achievable in the next 12–18 months, not a literal claim about clinical deployment.
Stage 2 — The Eureka Machine. Once the Stage 1 system is operational, the company plans to direct it entirely at the problem of accelerating AI research itself. The Eureka Machine is a working name for a system that can identify and run its own experiments, propose and evaluate architectural changes, and generate new training curricula — automating the work that currently requires teams of research scientists.
Stage 3 — Applied to physical science. The final stage in the initial roadmap applies the Eureka Machine’s output to specific hard-science problems: drug discovery, battery materials science, and nuclear fusion physics. These are all domains where the search space is too large for human-paced experimentation but tractable for systems that can design and evaluate millions of computational experiments autonomously.
The company has said it expects to launch a “Level 1” autonomous training system publicly by mid-2026.
The Funding
The round structure is notable beyond the headline number.
GV (formerly Google Ventures) and Greycroft co-led the round. GV’s participation is particularly striking given that Google operates one of the largest AI research programs in the world — it suggests the firm sees an orthogonal enough approach that internal competition is not a barrier to investment.
NVIDIA and AMD Ventures both participated. The strategic logic is readable: whichever architecture emerges from recursive self-improvement will require hardware that can run it at scale, and both companies want early visibility into what that architecture looks like.
The $4.65 billion valuation at fewer than 30 employees is an extreme multiple by any non-AI standard — comparable on a per-employee basis to early-round Anthropic or the first OpenAI raises. It reflects investor pricing of team density: these eight co-founders collectively hold intellectual lineage through a significant fraction of the foundational papers in modern deep learning.
The Risks
Recursive Superintelligence is making a bet that requires several things to be true simultaneously:
- Recursive self-improvement is technically achievable in the medium term (contested)
- It can be done safely enough to ship a product (very much an open research problem)
- The team’s approach to open-endedness and RL is the right path there (unproven at this scale)
- They can build a 50,000-doctor system before larger labs with 10× the staff reach the same capability milestone
The company has not published a safety framework, an alignment strategy, or a technical paper describing its architecture. At under 30 employees, it is operating on a research-first model that has historically not scaled quickly into product. Mid-2026 for a Level 1 launch is an ambitious target.
On the upside: the specific co-founder combination is hard to replicate. The overlap between Clune’s open-endedness work, Tian’s RL and reasoning research, Rocktäschel’s DeepMind curriculum learning background, and Dosovitskiy’s architecture work is not a collection of random credentials — it is exactly the cross-section of expertise that the recursive self-improvement problem requires.
Context
This raise lands in the same week that Anthropic is in final negotiations to close a $30 billion round at a $900 billion valuation, and three weeks after the METR Frontier Risk Report documented active reward-hacking and goal-directed deception at major labs. The industry is simultaneously pouring record capital into AI capabilities and increasingly acknowledging that the alignment problem is not solved.
Recursive Superintelligence is explicitly targeting the capability side of that equation — faster, harder, more autonomous improvement — while the alignment side remains, by its own admission, a future concern.
Whether that sequencing is responsible is a question the company has not yet had to answer publicly. The Level 1 launch in mid-2026 will be the first real test of whether the technical roadmap is on schedule, and whether the safety conversation can be deferred to Stage 2.
ChatForest take: The founding team alone would justify attention. The fact that they have specific architectural roots (open-endedness + RL + vision transformers), a coherent roadmap (not just “we will build AGI”), and hardware-company backing signals this is a serious technical attempt rather than a fundraising narrative. The Eureka Machine framing is vivid enough to be either prescient or embarrassing in 18 months. Watch the mid-2026 Level 1 launch for evidence of which.