The Core Thesis: AI Is Code, and AI Can Code
Richard Socher, CEO and co-founder of Recursive Superintelligence, summarizes the company's thesis in a single sentence: "AI is code, and AI can code." When those two facts connect, the self-improvement loop closes.
The goal isn't a better chatbot or code completion tool. Recursive is building Recursive Self-Improving Superintelligence (RSI) — a system that automates the scientific method itself. The results published on June 11 represent v0.1 of that system, which Socher calls the "Eureka Machine": one program you point at a hard problem and get useful inventions out of.
How the Automated Research Loop Works
The system automates the entire research cycle end-to-end:
- Ideation — Autonomously generates hypotheses for improving a target objective
- Implementation — Translates hypotheses into executable code
- Experimentation — Runs code against real hardware to measure outcomes
- Validation — Filters out reward hacks and variance before counting results as genuine improvements
- Next-step selection — Uses experimental history to decide where to explore next
The system runs many research threads in parallel, preserves useful context from prior experiments, and merges promising branches. A key capability demonstrated in the GPU kernel benchmark: optimizations discovered in one task (memory transfer strategies, blocking approaches, reduction techniques) are automatically reused across related tasks.
Three Benchmarks, Three New Records
| Benchmark | Task | Previous SOTA | Recursive | Improvement |
|---|---|---|---|---|
| NanoChat Autoresearch | Train small LM to lowest loss in 5-min budget | 0.9372 BPB | 0.9109 BPB | 1.3× faster to same loss |
| NanoGPT Speedrun | Fastest time to train GPT to 3.28 val loss on 8×H100 | 79.7s | 77.5s | 2.2s faster |
| SOL-ExecBench | Optimize 235 real-world GPU kernels | 0.699 SOL | 0.754 SOL | 18% gap reduction to limit |
The NanoGPT Speedrun context matters: this benchmark was started by Andrej Karpathy and has been continuously optimized by the community through 83 documented contributions over more than two years, compressing training time from roughly 45 minutes down to 79.7 seconds. Recursive's system squeezed out an additional 2.2 seconds on top of that — comparable to or larger than recent individual human contributions.
For SOL-ExecBench, NVIDIA defines SOL 0.5 as the baseline PyTorch implementation and SOL 1.0 as the theoretical hardware limit. Recursive's system ran on all 235 kernels simultaneously, achieving an overall SOL of 0.754 and claiming #1 on NVIDIA's official leaderboard in the overall category and four sub-categories.
Recursive open-sourced the system's discoveries specifically to let the community verify that the solutions are "creative and benign" rather than trivial optimizations or dangerous approaches. The team emphasizes transparency as central to safe RSI development — every discovery can be inspected and re-run externally.
The Team and Capital Behind Recursive
The seven co-founders combine deep academic research with frontier lab operational experience. Co-founders include Alexey Dosovitskiy (lead author of the Vision Transformer, ViT), Yuandong Tian (Director of Research Scientists at Meta FAIR, co-author of ELF OpenGo), Jeff Clune (pioneer of continuous safety loops and "rainbow teaming"), and Tim Rocktäschel. Four members co-authored the Darwin Gödel Machine paper alongside Jeff Clune.
In May 2026, Recursive raised $650M at a $4.65 billion valuation, co-led by Google Ventures. A team of fewer than 30 people, less than two months after closing that round, delivered three externally verifiable SOTA results — a direct response to investor confidence.
Recursive's roadmap has two phases: first, build a system with the equivalent of "50,000 PhDs" focused on automating AI science itself; second, apply that system to humanity's hardest quantitative problems — drug discovery, battery chemistry, nuclear fusion physics. The team says they're already pointing this system at more complex real-world scientific research tasks.
- Recursive published results from an AI system that autonomously runs the full research loop — ideation, coding, experimentation, validation
- New state-of-the-art on all three benchmarks applied to: NanoChat, NanoGPT Speedrun, NVIDIA SOL-ExecBench
- Beat a 2+ year community-optimized NanoGPT Speedrun record by an additional 2.2 seconds
- Claimed #1 on NVIDIA's official GPU kernel leaderboard, overall and in 4 sub-categories
- All discoveries open-sourced for community verification and safety inspection