Richard Socher has launched Recursive Superintelligence with $650 million in backing and an unusually ambitious brief: build self-improving AI systems that can spot their own weaknesses, redesign themselves, and help automate the whole research loop. That puts the startup squarely in the long-running race to make AI improve AI, a goal that sounds elegant in slides and deeply annoying in practice.
Socher, best known as the founder of You.com and a veteran AI researcher, says the real target is not just using models to write code or speed up experiments. He wants the entire scientific cycle automated: hypothesis generation, implementation, and verification, first in AI research and later, potentially, in other fields. That is a much taller order than the industry’s usual ”AI helps with work” pitch, and it places Recursive Superintelligence in the same conceptual neighborhood as other labs chasing agentic systems and self-improvement loops.
What Recursive Superintelligence is trying to build
The company’s core idea is ”open-ended evolution” for AI: a system that keeps adapting, competing, and finding new ways to get better instead of being locked into a fixed task list. In practice, that means models that can test each other, break each other’s defenses, and iterate through millions of rounds of competition. It is a familiar safety-testing trick in the AI world, but Recursive Superintelligence is betting it can be turned into a general engine for progress.
That ambition has a precedent. Major AI labs have already leaned on adversarial setups, self-play, and automated evals to harden models, but Socher is pushing the idea further: not just better benchmarks, but a system that learns how to improve the science itself. If it works, the company would be trying to turn research into a machine-readable loop rather than a human-led process with spreadsheets and caffeine.
The team brings deep lab experience
Recursive Superintelligence is also assembling a team that looks like it was built to impress investors and worry competitors. Former researchers from Google DeepMind and OpenAI are on board, including Tim Rocktäschel, known for work on self-learning and open-world systems, and Josh Tobin, who helped develop Codex and OpenAI research teams. That kind of talent matters because the hardest part of self-improving AI is not the slogan; it is making the loop stable enough to trust.
Socher is also making a point of not acting like this is only a research museum. The company wants products, too, and he says the first commercial release could arrive ”within quarters, not years.” That timeline is aggressive, but it is also a signal: the startup wants to sell something before the rest of the industry turns the same idea into yet another white paper.
Compute could become the real bottleneck
Long term, Socher thinks the limiting factor for a genuinely self-improving AI will not be human labor but computing power. That is a neat reversal of the usual fear that AI will replace researchers; in this version, the scarce resource becomes the amount of silicon you can throw at science. The bigger question, then, is not whether machines can help discover new things, but which scientific and medical problems humanity decides are worth paying for.
That debate is already getting more expensive. As model training and agentic systems scale up, the companies that can secure serious compute are the ones that get to define what ”progress” means. Recursive Superintelligence is making a bold bet that the next frontier is not a better chatbot, but a research engine that keeps rewriting itself until humans decide where to point it.