Recursive bags $650M to build self-improving AI

Most frontier AI labs are still running on a very human-heavy playbook: brilliant researchers design an architecture, feed it data, run an eval, and manually tweak the knobs.

Recursive wants to take humans out of that loop.

Operating as Recursive SI, the stealth startup just launched into the open with a massive $650 million funding round at a $4.65 billion valuation. The oversubscribed round was led by GV and Greycroft, with strategic backing from both AMD Ventures and NVIDIA.

The pedigree behind the round

Recursive is keeping its headcount lean, with a team of around 25 people split between San Francisco and London. But the talent roster reads like an all-star lineup of the modern deep learning boom.

The team is composed of former research leaders from OpenAI, Google DeepMind, Google Brain, Meta AI, Salesforce AI, and Uber AI. Its founders have previously built and sold startups to Big Tech giants and launched two separate unicorns.

Technically, they are not pitching another standard LLM wrapper or generic foundation model. Their stated expertise sits in open-ended algorithms, quality diversity algorithms, self-improving coding agents, automated red-teaming, world models, and building fully automated AI scientists.

Basically, instead of asking engineers to spend months tuning training runs, Recursive is building an AI that designs, runs, and evaluates its own experiments.

Darwinian evolution for neural networks

The company’s core thesis is inspired by evolutionary biology. Human intelligence was not hand-designed; it emerged through open-ended natural selection and cultural accumulation that continually built on earlier breakthroughs.

Recursive argues that the history of machine learning shows a constant trend: as compute scales up, hand-crafted human solutions get replaced by automated machine-driven discovery.

Their immediate roadmap focuses on applying this loop to AI science itself. Once they have a system that can reliably improve AI architectures and training regimens, they plan to deploy the same playbook across other scientific disciplines. They claim this recursive loop will be the fastest path to superintelligence, promising that automated safety checks and red-teaming will be baked into every step.

Why it matters

For everyday developers and end users, there is nothing to test or integrate right now. Recursive is operating strictly at the frontier research layer, meaning you will not be switching API keys anytime soon.

For the wider AI industry, this massive bet marks a philosophical shift. With traditional pre-training scaling laws facing compute walls and data exhaustion, the frontier is aggressively pivoting toward self-improving systems and autonomous scientific reasoning. Seeing Nvidia and AMD jointly back the round suggests the chip heavyweights believe recursive learning will require enormous compute for continuous inference and search loops.

We think targeting AI research as the first testbed is a smart tactical choice. AI development exists entirely in software, meaning synthetic data generation, code execution, and benchmark grading can provide instant reward signals without relying on messy real-world labs.

However, recursive self-improvement has always carried theoretical landmines. Without strict guardrails, self-modifying agents run the risk of feedback loops that reinforce narrow biases or cause sudden model degradation. Stepping onto the stage with a $4.65 billion valuation puts immense pressure on a 25-person shop to prove that an AI can actually do frontier research better than the humans who built it.