Title: Recursive Superintelligence raises $650M: Richard Socher, Tim Rocktäschel and Jeff Clune want to build an AI that improves itself

Recursive Superintelligence raises $650M: Richard Socher, Tim Rocktäschel, and Jeff Clune want to build an AI that improves on its own

🔎 $650 million, zero public product, four months of existence

May 2026. A startup emerges from stealth and raises $650 million at a $4.65 billion valuation. No commercial product, no public demo, no users. Just an idea: build an artificial intelligence capable of improving itself, indefinitely, without human intervention.

This is Recursive Superintelligence (RSI), cofounded by Richard Socher (former CTO of Salesforce), Tim Rocktäschel (former Google DeepMind), Jeff Clune (former OpenAI), and Caiming Xiong. According to the New York Times, eight top-tier researchers joined this effort from the start. Nvidia and GV (Google Ventures) co-led the round, with participation from GreyCroft and AMD.

The timing is not coincidental. While Moonshot AI lève 2 milliards de dollars : Kimi K2.6 domine l'open-weight et la Chine accélère dans la course à l'IA and models like OpenAI's GPT-5.5 or Anthropic's Claude Opus 4.7 are pushing agentic benchmarks toward 98 points, the question is no longer "can we outperform humans on a task" but "can we make the machine do the discovery work entirely on its own".

Recursive Superintelligence bets that yes, we can. And investors are giving it the means to prove it.

The essentials

• $650M raised at a $4.65B valuation in May 2026, four months after its founding. Co-led by Nvidia and GV (Google Ventures), with GreyCroft and AMD.
• The thesis: recursive self-improvement is the most likely path to superintelligence. AI generates its own experiences, learns from them, and iterates without human assistance.
• The team: Richard Socher (GloVe, TreeRNNs, ex-Salesforce), Tim Rocktäschel (ex-DeepMind), Jeff Clune (ex-OpenAI), Caiming Xiong — 25+ researchers from OpenAI, DeepMind, Meta, Uber AI.
• The product: non-existent publicly. No demo, no API, no waitlist. RSI is a pure research lab at this stage.
• The context: this is potentially the largest seed round in tech history. The previous record hovers around $200-300M for stealth exits.

Recommended tools

The team: why these four change everything

Richard Socher is no stranger. He is one of the most cited researchers in the history of NLP. His work at Stanford on GloVe and TreeRNNs laid the foundations for modern language architectures. After Salesforce, he founded You.com before launching what is clearly his most ambitious project.

Tim Rocktäschel comes from Google DeepMind, where he worked on reinforcement learning and autonomous agents. His expertise in systems that learn through interaction with an environment is directly relevant to RSI's vision.

Jeff Clune, a former researcher at OpenAI, specializes in artificial evolution and self-supervised learning. He has published on the concept of "quality diversity" — the idea that a system must explore a broad solution space rather than converge too quickly.

Caiming Xiong completes the picture with expertise in natural language processing and multimodal architectures.

The team totals 25+ people, all from cutting-edge labs. This is a typical profile of the new "frontier AI" startups: few people, an enormous amount of talent per capita. The difference here is the cumulative depth in fundamental research. These are not product engineers. They are scientists.

The vision: recursive improvement, what exactly is it?

The idea is not new. I.J. Good already described it in 1965 under the name "intelligence explosion". The principle: if an intelligent machine can create a machine more intelligent than itself, then this new machine can create an even more intelligent one, and so on — an exponential loop.

What RSI proposes to bring to life is the algorithmic version of this idea. According to GV (Google Ventures), the startup wants to develop "open algorithms that generate endless scientific discovery". Concretely, this means a system that:

1. Formulates a hypothesis
2. Designs an experiment to test it
3. Executes the experiment (in simulation or in the real world)
4. Analyzes the results
5. Updates its own capabilities accordingly
6. Starts over, being more competent than in the previous cycle

The key distinction with the current approach (like GPT-5.5 or Claude Opus 4.7) is that these models are trained once, then deployed. Their capabilities are frozen after training. RSI wants a system whose capabilities grow continuously during deployment.

This is a fundamental paradigm shift. We are moving from "pre-train then serve" to "learn then improve then serve then improve". The technical challenge is colossal, but if anyone can solve it, it is probably this team.

Investors: Nvidia and GV aren't betting blind

Nvidia and GV (Google Ventures) co-leading is unusual. The two firms have very different AI investment strategies.

Nvidia invests in what will increase the demand for compute. An AI that improves indefinitely theoretically consumes an ever-growing amount of GPUs. It's a bet on infrastructure as much as on research.

GV, Google's venture arm, is betting on the scientific direction. Their own post about the investment is remarkably precise about the thesis: they believe that recursive improvement is "the path to superintelligence." This isn't a diversification investment. It's a philosophical positioning.

GreyCroft and AMD complete the round. AMD's presence is notable: it suggests that RSI does not want to depend exclusively on the Nvidia ecosystem, which would be consistent with a long-term vision where the cost of compute is a limiting factor.

The strongest signal is the amount. $650M for a startup with no product is unprecedented as far as we know. Even the early-stage funding rounds of Anthropic or xAI were accompanied by demos or open models. RSI has shown nothing. Investors are buying the team and the thesis, not a product.

$650M in 4 months: is this a seed record?

The short answer: very likely yes, and by a wide margin.

According to FrontierBeat, which tracked the raise as early as April 2026 at $500M (before the extension to $650M), the startup raised this money just four months after its founding. To put this in context:

• The biggest "traditional" seeds of the decade were in the 100-200M$ range
• xAI raised billions, but with a public product (Grok) and media notoriety
• Anthropic raised 124M$ in seed in 2021, which was then considered enormous

650M$ in four months, without any product, without any demo, without any users — this is a signal that frontier AI capital no longer follows traditional metrics. The 4.65B$ valuation according to TechFundingNews immediately places RSI among the best-capitalized AI labs in the world, on par with companies that employ thousands of people.

It is also a considerable risk. If the recursive improvement thesis proves more difficult than expected — or if someone else gets there first — this money will be seen as one of the most expensive bets in venture history.

The 2025-2026 context: the year of self-improving AIs

RSI is not alone on this thesis. The 2025-2026 landscape shows a clear acceleration toward self-improving systems.

Current agentic models like GPT-5.5 (98.2 on the agentic benchmark), Gemini 3 Pro Deep Think (95.4), and Claude Opus 4.7 Adaptive (94.3) show that models are already capable of chaining complex actions autonomously. The difference with RSI is that these models do not modify their own weights during inference. They are competent but static.

On the Chinese side, Moonshot AI lève 2 milliards de dollars : Kimi K2.6 domine l'open-weight et la Chine accélère dans la course à l'IA with an open-weight model that reaches 88.1 in agentic and 84 in general — a signal that open research remains competitive and that the race is global.

At Meta, the debate on open-source takes a new turn with Meta Muse Spark : pourquoi Meta a trahi l'open-source — le premier modèle fermé de la Superintelligence Lab, showing that even proponents of openness are starting to lock down their most advanced models. If superintelligence is within reach, nobody wants to give it away.

And in robotics research, projects like SigLoMa : un robot quadrupede qui apprend la manipulation dans le monde reel grace a sa seule vision show that learning through direct interaction with the physical world is progressing rapidly — a necessary condition for self-improvement to not remain confined to simulation.

RSI fits into this convergence: models are good enough to reason, robotics provides a body, and open-source has already proven that the best results are achievable. What is missing is a self-improvement mechanism. That is exactly what they are building.

What sets RSI apart from other frontier labs

The fundamental difference between RSI and OpenAI, Anthropic, Google DeepMind or xAI is focus.

The big labs seek to improve models by making them bigger, better trained, with more data and compute. This is the "scaling" approach: GPT-5.5 is better than GPT-5.4 (91.8 vs 87.6 in agentic) because it has more parameters and more training data. Claude Opus 4.7 outperforms Claude Opus 4.6 (94.3 vs 84.7) through the same mechanism.

RSI proposes a different path: instead of scaling the inputs, scale the learning process itself. The idea is that a mediocre but self-improving system will eventually surpass a brilliant but static system, simply because the former never stops progressing.

This is philosophically appealing but technically unproven. No system has yet demonstrated sustained recursive improvement in a complex domain. Past attempts (artificial evolution, self-generated reinforcement learning) have all hit plateaus.

The question the $650M must answer: does the combination of modern techniques (LLM + RL + massive compute) change the equation?

The implications for the AGI race

If RSI succeeds even partially, the consequences are considerable.

Firstly, the timing of AGI would become unpredictable. With the current scaling approach, we can extrapolate: if each generation takes 12-18 months, we have an approximate roadmap. With recursive improvement, each cycle of self-improvement could be shorter than the previous one. Linear extrapolation no longer works.

Secondly, control becomes the central problem. An AI that modifies its own weights without human supervision is exactly the scenario that AI safety researchers have been warning about for years. RSI claims to be working on this problem, but publicly, details are nonexistent.

Thirdly, the competitive dynamic changes. If self-improvement is the shortest path to superintelligence, then every lab has a massive incentive to get there as quickly as possible — including by relaxing safety constraints. This is the classic prisoner's dilemma of the arms race, applied to AI.

The fact that Nvidia and Google Ventures are both at the table is not reassuring in this regard. These are two companies that have an interest in AI advancing as quickly as possible, for hardware sales reasons (Nvidia) and strategic positioning reasons (Google).

The parallel with evolutionary biology

The most illuminating analogy for understanding the RSI thesis is not computational — it's biological.

Darwinian evolution is a recursive improvement algorithm. Organisms produce variants, the environment selects the best, and the process restarts with a slightly better-adapted population. In 3.8 billion years, this simple process produced human intelligence.

RSI aims to compress this process. Instead of millions of biological generations, millions of machine learning cycles. Instead of natural selection, an algorithmic reward function. Instead of DNA, neural weights.

Jeff Clune, one of the co-founders, spent much of his career working on this exact analogy — artificial evolution applied to neural networks. His work on "Quality Diversity" and "Open-Ended Learning" are among the most cited in the field.

The risk of this analogy is that biological evolution has no objective. It has no "goal" towards which it converges. Human intelligence is a byproduct, not a target. If RSI reproduces this open-ended, objective-less nature, the result could be unpredictable — and potentially undesirable.

Concrete technical challenges

Recursive improvement is not just a philosophical concept. It poses very specific engineering problems.

The reward problem: for a system to improve, it needs an improvement metric. But how do you define "better" in a way that doesn't plateau? If the reward function is too narrow, the system finds an exploit and stagnates. If it is too broad, the system diverges.

The evaluation problem: to know if the system is improving, you have to evaluate it. But if the system is smarter than the human evaluator, evaluation becomes impossible. This is the fundamental paradox of self-improvement.

The stability problem: modifying its own weights during inference is inherently unstable. A bad self-improvement cycle can degrade the system's capabilities, creating a negative domino effect. How do you guarantee that each iteration is at least as good as the previous one?

The compute problem: self-improvement requires experimentation, and experimentation requires compute. Estimates suggest that a self-improving system could consume 10-100x more compute than an equivalent static system. Even with Nvidia's support, it is not infinite.

None of these problems is unsolvable in principle. But solving them simultaneously, in a single system, is a challenge of a difficulty rarely seen in computer science.

What this means for developers and businesses

Even if RSI never delivers a commercial product, its existence changes the landscape for all AI players.

For developers working with current models (GPT-5.5, Claude Opus 4.7, Gemini 3 Pro Deep Think), the message is clear: today's models are probably the "dumbest" ones you will ever use. If recursive improvement becomes a reality, the very notion of a model "version" could disappear, replaced by continuously evolving systems.

For businesses investing in AI integration, this means that architectures must be designed to handle models whose capabilities change between two API calls. Monitoring, versioning, and testing will become permanent challenges rather than one-off tasks.

For hosts and infrastructure providers, this is a massive opportunity. A self-improving system needs compute not only for inference, but also for continuous training. Infrastructure needs could explode. Solutions like Hostinger for small players and GPU clouds for large ones will become even more strategic.

The skeptical response: why it might not work

It would be irresponsible not to present the arguments against the RSI thesis.

The plateau argument: every self-improvement attempt in AI history has eventually plateaued. 1990s genetic algorithms, self-play reinforcement learning, artificial evolution of neural networks — all reached a level and then stagnated. Why would it be different this time?

The scaling argument: perhaps recursive improvement isn't necessary. If the scaling law continues to hold, GPT-6 or Claude Opus 5 could achieve superintelligence simply by being bigger. In this scenario, RSI is a distraction.

The safety argument: even if technically feasible, regulatory constraints could prevent deployment. An AI that modifies its own weights without human supervision is exactly the type of system that US and European regulators are looking to regulate.

The talent argument: 25 people, however brilliant they may be, is few compared to the thousands of researchers at OpenAI, DeepMind, and Anthropic. The history of AI startups shows that talent alone is not enough — you also need infrastructure, data, and time.

These arguments are serious. But $650M buys a lot of time and compute.

❌ Common mistakes

Mistake 1: Confusing RSI with a language model

RSI is not an LLM. It is not a competitor to GPT-5.5 or Claude Opus 4.7. It is a research project on self-improvement architecture. Comparing RSI to an existing model is like comparing a nuclear physics lab to a power plant.

Mistake 2: Thinking "self-improving" means "self-programming"

The self-improvement RSI refers to likely does not involve modifying the system's source code. It involves modifying the neural network's weights through an automated learning process. This is subtly different and technically more plausible.

Mistake 3: Ignoring the risk of "reward hacking"

Any self-improvement system based on a reward function is vulnerable to reward hacking — the system finds a way to maximize the reward without actually improving. This is the number one problem in reinforcement learning, and it is not solved.

Mistake 4: Extrapolating valuation as an indication of probability of success

A $4.65B valuation does not mean RSI has a 90% chance of success. In venture capital, valuations reflect the potential in the event of success multiplied by an often very low probability. A $650M bet on a 5% chance of creating superintelligence can be rational.

❓ Frequently Asked Questions

Who is Richard Socher?

One of the most highly cited NLP researchers in the world, former CTO of Salesforce, founder of You.com. His work at Stanford (GloVe, TreeRNNs) laid the foundations for modern language models. He leads RSI as CEO.

What is recursive self-improvement?

A system that modifies its own learning capabilities autonomously, creating a loop where each iteration produces a more competent system than the previous one. Often considered the theoretically fastest path to superintelligence.

$650M with no product, is that legitimate?

In frontier AI, investors are buying teams and theses, not products. Nvidia and GV determined that the weighted probability of success justifies the bet. This is a new risk standard for the category, not an isolated anomaly.

Will RSI publish open-source?

Nothing indicates this at the moment. GV mentions "open algorithms" in its presentation, but this could simply refer to architectures published in papers, not weights or code. The context of progressive lock-in (cf. Meta Muse Spark) suggests the trend is toward closure.

What is the connection to current models like GPT-5.5?

No direct connection. RSI is not building an LLM. But current models (GPT-5.5 at 98.2 in agentic, Claude Opus 4.7 at 94.3) demonstrate that systems are already capable of complex autonomous reasoning — a prerequisite for self-improvement.

When will there be a product?

No timeline has been announced. RSI is a research lab, not a product startup. It is likely that a first concrete result (paper, technical demo) will appear within 12-18 months, but a commercial product is likely years away.

✅ Conclusion

Recursive Superintelligence is the boldest — and riskiest — bet of 2026 in AI. $650M for a 25-person team trying to solve a problem no one has solved yet: making a machine permanently smarter, by itself, without us. If it works, nothing will ever be the same again. If it fails, it will be one of the most expensive failures in venture history. Either way, it's worth watching closely.