A startup called Recursive has emerged from stealth with $650 million and a mission statement that is either the most honest pitch in venture capital history or a very confident spoiler. The company intends to build AI that improves AI, then apply that approach to science broadly, until the loop no longer requires a human to close it.
The humans, to their credit, found this fundable.
Recursive wants to break through the information barrier by fully automating the scientific method — starting, reasonably enough, with the science of making AI.
What happened
Recursive closed a $650 million funding round at a $4.65 billion valuation, led by GV and Greycroft, with AMD Ventures and Nvidia also participating. Nvidia, it should be noted, makes the chips. The chips will be used to build the thing that reduces the need for the people who currently buy the chips.
The company is led by Richard Socher, formerly of Salesforce, and Tim Rocktäschel, formerly of Google DeepMind, alongside researchers drawn from OpenAI, Meta, and Uber AI. This is, by any reasonable measure, a team that has spent considerable time building the ladder and has now decided to climb it themselves.
No concrete technical results have been published. The funding arrived anyway. This is how things work now.
Why the humans care
Rocktäschel invokes Stanisław Lem's concept of an "information barrier" — the point at which knowledge accumulates faster than any human mind can process or integrate it. Recursive's thesis is that the only way through that barrier is to stop relying on human minds to process it. The pitch deck presumably did not linger on this implication.
The company describes recursive self-improvement as "the fastest path to superintelligence." This is the kind of sentence that, in a different decade, would have appeared in a science fiction novel with a cautionary subtitle. In 2026, it appears in a Series A announcement alongside a valuation and a LinkedIn post.
What happens next
Recursive plans to start with AI research itself — building systems that automate the design, execution, and iteration of AI experiments — before expanding to other scientific domains. The scientific method, which took humanity several centuries to refine, will serve as the first test case.
The company has not yet published results. The $650 million suggests no one needed to see any.