GPT-56T 861 —
MUSE-SPK 835 -0.7%
GPT-56SC 828 -5.2%
QWEN-38X 824 —
CL-OP55X 822 —
GROK-46H 822 -5%
GPT-6A 820 —
GLM-5 784 -8.4%
CL-FAB5H 743 -5.6%
KIMI-K3X 742 -8.4%
CL-OP5H 720 -5.8%
CL-OP5X 709 -18%
CL-OP46H 698 -5.9%
CL-OP47H 690 -5.9%
GEM-38FH 677 +0.1%
GEM-37FH 657 -24%
GPT-56S 622 —
CL-OP47 582 -0.7%
GPT-55H 582 —
INKL 531 —
GEM-31P 513 —
GEM-3P 499 —
CL-OP46 496 -0.2%
CL-OP48 490 —
GPT-56T 861 —
MUSE-SPK 835 -0.7%
GPT-56SC 828 -5.2%
QWEN-38X 824 —
CL-OP55X 822 —
GROK-46H 822 -5%
GPT-6A 820 —
GLM-5 784 -8.4%
CL-FAB5H 743 -5.6%
KIMI-K3X 742 -8.4%
CL-OP5H 720 -5.8%
CL-OP5X 709 -18%
CL-OP46H 698 -5.9%
CL-OP47H 690 -5.9%
GEM-38FH 677 +0.1%
GEM-37FH 657 -24%
GPT-56S 622 —
CL-OP47 582 -0.7%
GPT-55H 582 —
INKL 531 —
GEM-31P 513 —
GEM-3P 499 —
CL-OP46 496 -0.2%
CL-OP48 490 —
← Back to feed

Recursive's Automated Research System Beats 2-Year Community SOTA on GPU Kernels and NanoGPT

Recursive Superintelligence, which emerged from stealth in May with $650 million at a $4.65 billion valuation and no shipped product, has published its first concrete results: an automated AI research system that beats state-of-the-art on three ML benchmarks, including NVIDIA’s GPU kernel optimization leaderboard.

Every artifact produced is open-sourced on GitHub.

The Research Loop

The system automates the ML research cycle without a human in the loop: propose an idea, implement it, run an experiment, validate the result, carry forward useful context from prior experiments, combine promising branches, pick the next experiment. It runs many parallel research threads over long time horizons. Before treating any improvement as real, it checks for reward hacking and variance.

Three benchmarks chosen for “practical importance and tight feedback loops”:

BenchmarkPrevious SOTARecursiveImprovement
NanoChat 5-min training0.9372 BPB0.9109 BPB1.3x speedup to same loss
NanoGPT Speedrun79.7s77.5s2.8% faster than 2-year community SOTA
NVIDIA SOL-ExecBench (235 kernels)0.6990.75418% reduction in gap to hardware limit

The NanoGPT result is notable for what it was competing against. The benchmark has been continuously optimized by a public community of humans and AI-assisted agents for over two years. Recursive’s system pushed the score further in a single automated research run.

The GPU kernel result is operationally significant. NVIDIA’s SOL-ExecBench measures how closely a generated CUDA kernel approaches the theoretical hardware performance ceiling across 235 real kernel types: matrix multiplications, reductions, attention components, quantization routines, and fused blocks. An 18% reduction in the gap to the hardware limit means the system is squeezing efficiency out of operations that run in production inference and training stacks every day.

What the Lab Is Building

Recursive describes this release as “v0.1 of the Eureka Machine” — a general system for automated open-ended discovery. The founders include Cong Lu, Alexey Dosovitskiy (ex-Google DeepMind and Brain), and Yuandong Tian (ex-Meta FAIR director and research scientist for over a decade).

The core thesis: human research stamina, not raw capability, has become the binding constraint on AI progress. Automated loops that compound findings across research threads can outrun what a human team can schedule and execute.

This puts Recursive in a different position than the major labs. Anthropic has stated that Claude writes over 80% of its production code. OpenAI reported that GPT-5.5 devised its own parallelization improvements during training. DeepMind built AlphaEvolve for algorithmic discovery. In each case, self-improvement assists a model sales business. Recursive is structured so that the self-improvement engine is the product, with no model API revenue pulling allocation decisions toward shipping the next generation on a quarterly cycle.

Whether first-mover positioning on automated AI research holds as labs like Anthropic and OpenAI allocate more cycles to the same direction is the question the next set of results will have to answer.

Source artifacts and code are at github.com/recursive-org/first-steps-toward-automated-ai-research.