$650M Backs Richard Socher's Recursively Self-Improving AI

On May 13, 2026, Recursive Superintelligence emerged from stealth with a $650 million funding round to build AI systems capable of recursive self-improvement. Founded by former Salesforce Chief Scientist Richard Socher, the four-month-old startup reached a $4.65 billion valuation. The capital is earmarked for securing the massive GPU infrastructure required to run millions of autonomous training experiments.

Architecture of Self-Improvement

The technical strategy centers on “open-endedness,” a design paradigm inspired by biological co-evolution. Instead of relying on human engineers to manually design architectures, the system is built to identify its own architectural flaws and rewrite its underlying code. Co-founder Tim Rocktäschel refers to this as breaking the Information Barrier, a threshold where technical knowledge accumulates faster than humans can process it.

The system is designed to automate the generation, implementation, and technical validation of research ideas. By creating autonomous loops that hypothesize and test architectural changes, the startup aims to shift the primary bottleneck of artificial intelligence from human intuition to raw compute availability. By shifting this bottleneck to automated design space exploration, the system learns how to learn.

Compute and Strategic Backing

Automating the entire scientific method requires extraordinary computing scale. The oversubscribed round was led by GV and Greycroft, with strategic participation from Nvidia, AMD Ventures, and Salesforce Ventures.

The dual backing from both major hardware vendors underscores the extreme compute consumption inherent in running autonomous optimization loops. As you build AI agents, the cost of testing and iteration is a primary constraint. Recursive Superintelligence plans to absorb this cost at a multi-billion-dollar scale. Securing direct investment from Nvidia and AMD provides early access to the hardware necessary for this operation, as millions of parallel training experiments require infrastructure that exceeds the requirements of standard pretraining runs.

Research Leadership and Roadmap

Founder Richard Socher brings commercial deployment experience from his time at Salesforce and You.com. The founding team brings together 25 to 30 researchers from leading artificial intelligence labs:

• Tim Rocktäschel: Co-founder and Chief Scientist, formerly a principal scientist at Google DeepMind.
• Yuandong Tian: Co-founder, previously a research scientist director at Meta’s FAIR.
• Alexey Dosovitskiy: Lead researcher and co-author of the original Vision Transformer (ViT) paper.
• Josh Tobin, Jeff Clune, and Tim Shi: Co-founders with extensive engineering backgrounds from OpenAI.

Despite its heavily theoretical foundation, the company targets a mid-2026 public launch for its initial products. The immediate milestone is training a specialized system with the baseline capability of 50,000 domain experts to automate AI scientific research before expanding into other disciplines. If you are monitoring AI applications in production environments, managing the unpredictable trajectories of self-modifying code will require entirely new validation frameworks.

The shift from static model weights to continuously self-modifying architectures changes how compute budgets are allocated. Developers preparing for recursive systems should begin decoupling their application logic from fixed model versions, as underlying capabilities will iterate without human deployment cycles.