Tim Rocktäschel co-founds Recursive to build self-improving AI
The UCL AI professor joins a heavyweight founding crew to turn compute into accumulated knowledge with open-ended, automated scientific discovery.
By Ryan Merket ·
Why it matters
If Recursive can safely automate parts of the scientific method, founders and labs could turn compute into durable research assets instead of single-shot model runs.
Tim Rocktäschel (@_rockt) has co-founded Recursive, a new company aiming to automate the scientific method starting with AI research. In a launch post, he said the team in London and SF is building systems that can safely experiment on how to improve themselves.
The bet
Rocktäschel, a Professor of AI at UCL Centre for AI, framed the thesis plainly: "create AI that experiments on how to safely improve itself, turning compute into knowledge that accumulates in an open-ended process of endless, automated scientific discoveries" he wrote on X. The near-term focus is converting raw compute into reproducible insights rather than just bigger models, and doing so in a way that measures and contains risk as systems iterate on their own capabilities.
Who is building it
The founding roster goes well beyond one lab head. In a post on X, Christopher Manning listed the founders as Richard Socher, Rocktäschel, Jeff Clune, Tim Shi, Yuandong Tian, Caiming Xiong, Alexey Dosovitskiy, and Josh Tobin, adding that AIX Ventures backed the company from day one. Richard Socher also shared coverage and pointed readers to a New York Times writeup of the launch.
What Recursive says it will build
From Rocktäschel's post and the announcement thread, Recursive is positioning itself around three ideas:
- Self-improvement as a first-class capability: systems that can run experiments on themselves to get better.
- Safety baked into the loop: explicit attention to how to contain and evaluate self-modification.
- Knowledge that compounds: infrastructure that turns compute into an ever-growing body of findings, not just single-run results.
That focus puts Recursive in the lane of automating parts of research workflows that today depend on human iteration: proposing hypotheses, running controlled experiments, evaluating, and feeding results back into future trials. The hard part is doing this in an "open-ended" way without drifting into unsafe behavior or trivial self-optimizations.
Why now
Automation has long touched AI research through things like neural architecture search and large-scale ablations, but the founders here are explicitly framing a broader, sustained scientific process rather than one-off searches. With teams in London and San Francisco and a bench of researchers who have spent their careers at the frontier of machine learning, the company is betting that better tooling around experimentation will unlock qualitatively new capabilities.
Recursive has not published a public product page yet in the materials we saw, but the company emerged from stealth this week with a steady stream of posts from founders and supporters. As Rocktäschel put it, the goal is not just to build another model, but to "accumulate" knowledge through endless automated discovery. If they can show real, safe, reproducible gains from that loop, it could change how labs and startups alike do R&D.
Sources: Rocktäschel's launch post, Christopher Manning's founder list, Richard Socher's share of NYT coverage.