Recursive Superintelligence emerged from stealth with fewer than 30 staff, no product, and a $4.65 billion valuation and Richard Socher left a $1.5 billion company to build it. We think that decision deserves a proper kind of examination.
The idea of a machine that improves itself - that identifies its own weaknesses, designs its own fixes, and implements them without human intervention, then does it again, faster, in an accelerating loop that eventually outpaces every human researcher on earth, has been a fixture of computer science folklore since the 1960s.
For most of that time, it remained comfortably theoretical. The kind of thing you discussed in seminars and wrote papers about and filed quietly under 'fascinating but not yet.'
In May 2026, someone raised $650 million to build it in London.
Recursive Superintelligence emerged from stealth with a team of fewer than 30 people, no released product, offices in London and San Francisco, and a valuation of $4.65 billion. The round led by GV, Google's venture arm, and Greycroft, with participation from Nvidia and AMD, was described as heavily oversubscribed.
The founders are Richard Socher, formerly chief scientist at Salesforce, and Tim Rocktäschel, professor of AI at University College London and former principal scientist at Google DeepMind. Socher left You.com, a company he had built to a $1.5 billion valuation, to pursue this and Rocktäschel left one of the most prestigious AI research roles in the world.
They started recruiting researchers from OpenAI, Meta, and Google Brain, then they raised more money in a single round than most UK startups will see in their entire existence.
That's the part worth paying attention to.
The Idea That Attracted the Funding Behind Recursive Superintelligence
Most AI progress to date has been achieved by human researchers designing better models, better training methods, and better evaluation frameworks.
Humans decides what to try.
Humans assess whether it worked.
Humans designs the next iteration.
The process is brilliant, but it's also expensive, slow, and limited by how much work humans can do.
Recursive Superintelligence's central argument is that this bottleneck is the problem worth solving, by building AI systems that can replace the human researcher in the loop entirely. Systems that evaluate their own weaknesses, design improved training algorithms, select their own data, run their own experiments, and iterate without waiting for a human to review the results.
"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." — Recursive Superintelligence
The inspiration, as Socher has explained it, is biological evolution.
Natural selection produced eyes, language, and human intelligence through a process of endless iteration with no ceiling and no external designer.
Recursive Superintelligence wants to build the computational equivalent which is an open-ended algorithm that co-evolves with itself, finding solutions that no human researcher would have designed, at a speed no human team could match.
One practical example already in use across major AI labs which is rainbow teaming, a technique developed by Rocktäschel's team at DeepMind, where two AI systems co-evolve - one attempting to make the other produce harmful outputs, the other learning to resist.
The iterative adversarial process generates robustness that hand-designed safety methods cannot. That technique, Socher has noted, is now used at every major AI lab. It's also a preview of the broader methodology Recursive is pursuing at scale.
A Very British AI Story
Recursive Superintelligence is incorporated in London. Rocktäschel being a UCL professor, the company's ties to the UCL Centre for Artificial Intelligence and the broader Knowledge Quarter - the concentration of academic and research institutions around King's Cross - are deliberate rather than incidental.
The company joins Nscale, Wayve, Stability AI, and a growing cohort of frontier AI companies choosing London as their primary base. Among frontier AI companies emerging from London AI hubs, it stands out not just for its ambitious thesis but for the calibre of talent assembled to pursue it.
The combination of world-class academic AI research, a regulatory environment that is meaningfully more permissive than the EU on frontier AI development, and a talent pool drawing from UCL, Imperial, Oxford, and Cambridge creates conditions that the US cannot fully replicate and the EU is actively making harder.
"It is exceptional to see researchers of this calibre choosing London as the base for companies of such ambition." — Professor Geraint Rees, UCL Vice-Provost
The UK government has explicitly stated its intention to be an AI maker rather than a consumer. Recursive Superintelligence - $4.65 billion valuation, fewer than 30 staff, founded in 2025, already backed by Google and Nvidia - is probably the clearest single proof point that this ambition has substance behind it.
The emergence of such companies from the London AI ecosystem signals a structural shift in where cutting-edge AI research is being commercialised.
So What Happens Next
The honest position on Recursive Superintelligence is that this is a very large bet on a very hard problem.
Recursive self-improvement in AI has been theorised for decades but nobody has built it. The gap between a compelling research thesis and a working system is the same gap that has swallowed many well-funded, well-credentialled AI labs before this one.
What is different here and what the $650 million reflects, is the specific combination of people involved.
Socher has built and scaled AI products before.
Rocktäschel has spent his career working on precisely the technical problems Recursive is trying to solve. The team around them includes people who have contributed to the foundational research at every major lab. This is not a group that is new to the problem.
The Level 1 autonomous training system – the very first milestone - is targeted for public launch in mid-2026. That is a concrete, time-bound commitment from a company that has so far operated entirely in stealth.
We shall be watching very closely to see what it looks like when the curtain finally comes down.
Britain has a habit of producing foundational AI research and watching the commercialisation happen elsewhere.
Recursive Superintelligence is a direct bet against that exact pattern.
Whether it pays off is, by definition, unknown. Since the calibre of the people making the bet suggests the question is worth taking seriously.
