Recursive Superintelligence raises $650M to build self-improving AI models

Recursive Superintelligence Inc., a startup focused on building self-improving artificial intelligence, announced a $650 million funding round led by Alphabet Inc.’s GV fund and Greycroft, with participation from Nvidia Corp. and AMD’s venture arm. The investment values the company at $4.65 billion. Founded earlier this year by Richard Socher, former Chief Scientist at Salesforce and founder of You.com, Recursive now employs over 25 people across San Francisco and London. The company’s goal is to develop AI models that can autonomously improve their own code and infrastructure through automated scientific discovery. Unlike current neural networks, which lack full autonomy, Recursive’s AI will generate experiment ideas, test them, and validate results. The system will also optimize its training and inference processes, similar to how OpenAI’s GPT-5.5 improved token generation speeds by over 20% through dynamic parallelization. Recursive’s approach differs from rivals like Ineffable Intelligence, which uses reinforcement learning. The startup plans to start with AI research before expanding into physics, chemistry, and pre-clinical biology. Socher has described AI as a potential ‘new language’ for understanding complex systems, akin to calculus in physics. The funding round underscores growing competition in advanced AI development, with investors like Alphabet already using neural networks to design hardware like TPUs. Recursive will implement guardrails to prevent risky outputs while focusing on iterative self-improvement, positioning itself as a leader in next-generation AI systems.