Recursive

Co-founder and CEO Richard Socher earned his computer science PhD at StanfordStanford UniversityStanford University is a research university in California founded by Leland and Jane Stanford in 1885 and opened in 1891.Open the full entry in 2014. He co-founded MetaMindMetaMindMetaMind was a deep learning company co-founded by Richard Socher and Sven Strohband. It offered text and image analysis and was acquired by Salesforce in 2016.Open the full entry, serving as CEO and CTO, and became chief scientist and executive vice president at SalesforceSalesforceSalesforce is a U.S. enterprise software company headquartered in San Francisco, providing cloud applications, data platforms and AI agent tools centered on customer relationship management (CRM).Open the full entry after its 2016 acquisition of the company. He subsequently founded You.comYou.comYou.com is an AI search and infrastructure company founded by Richard Socher and Bryan McCann. It provides search, content extraction and research tools connecting AI models and agents with web information.Open the full entry and co-founded AIX VenturesAIX VenturesAIX Ventures is an early-stage AI venture firm co-founded by Richard Socher and Shaun Johnson in 2021, combining AI research, entrepreneurial and investment experience.Open the full entry, combining research, company-building and investment experience before Recursive.⁠[1]

Recursive Superintelligence emerged publicly on May 13, 2026. Socher’s announcement named seven fellow co-founders: Tim RocktäschelTim RocktäschelTim Rocktäschel is a German-born artificial intelligence researcher who is Professor of Artificial Intelligence at University College London and co-founder and chief scientist of Recursive Superintelligence; he previously led the Open-Endedness team at Google DeepMind.Open the full entry, Tim ShiTim ShiTim Shi (Tianlin Shi) is an artificial intelligence researcher and entrepreneur who worked on web agents in OpenAI’s early days, co-founded the conversational AI company Cresta in 2017 and served as its CTO, and later co-founded Recursive Superintelligence.Open the full entry, Josh TobinJosh TobinJosh Tobin is a machine learning researcher and entrepreneur who was a research scientist at OpenAI, co-founded the machine learning platform company Gantry, and is a member of the founding team of the AI research company Recursive Superintelligence.Open the full entry, Caiming XiongCaiming XiongCaiming Xiong is an AI researcher and co-founder of Recursive, formerly senior vice president of AI research at Salesforce.Open the full entry, Jeff CluneJeff CluneJeff Clune is an artificial intelligence researcher who is a professor of computer science at the University of British Columbia and a Canada CIFAR AI Chair at the Vector Institute; he works on open-ended algorithms and “AI-generating algorithms,” and in 2026 co-founded Recursive.Open the full entry, Yuandong TianYuandong TianYuandong Tian is an artificial intelligence researcher, science-fiction writer, and co-founder of Recursive Superintelligence. Formerly a research scientist director at Meta FAIR, he has worked on Go AI, long-context inference, efficient training, latent-space reasoning, and neural-network learning dynamics. His research includes ELF OpenGo, StreamingLLM, GaLore, and Coconut.Open the full entry and Alexey DosovitskiyAlexey DosovitskiyAlexey Dosovitskiy is a computer vision and machine learning researcher and a Recursive co-founder, with contributions to FlowNet, CARLA and Vision Transformer.Open the full entry. Their backgrounds span language models, computer vision, reinforcement learning, robotics and open-ended learning.⁠[2][3]

Investor GV disclosed $650 million in early financing at a $4.65 billion valuation. GV and Greycroft co-led the investment, with NVIDIANVIDIANVIDIA is an American semiconductor and computing technology company that provides graphics processing units, artificial intelligence infrastructure, networking interconnects, and computing software, and is listed on Nasdaq under the ticker NVDA.Open the full entry and AMDAMDAMD (Advanced Micro Devices, Inc.) is a publicly traded U.S. semiconductor company founded in 1969 and headquartered in Santa Clara, California; its stock trades on Nasdaq under the ticker AMD. The company develops CPUs, GPUs, FPGAs, adaptive systems-on-chip and related software, mainly for the data center, personal computer, gaming and embedded markets; its product brands include Ryzen, EPYC, Radeon and Instinct.Open the full entry participating. Recursive has offices in San Francisco and London. The valuation describes the company’s assessed value, rather than the cash raised.⁠[2]

Recursive aims to automate a research cycle in which AI proposes ideas, implements them, runs experiments and evaluates results to choose subsequent directions. The company draws inspiration from open-ended processes such as evolution and cultural accumulation. Its initial focus is AI research; expansion into other sciences is a longer-term aim.⁠[3][4]

In a Silicon Valley 101 interview released on June 5, co-founder Yuandong Tian described collaborating with Caiming Xiong on modeling, training and post-training, and agent harnesses.⁠[5]

Recursive’s updates index dates its initial automated-research results to June 11, 2026. The experiments covered NanoChat, NanoGPT Speedrun and SOL-ExecBench. The system pursues parallel research branches and uses experimental feedback to improve implementations across successive iterations.⁠[6][7]

For NanoChat, the company compared five-minute training runs on a single B200 GPU. After removing evaluation exploits, the baseline averaged 0.9372 bits per byte across ten random seeds, versus 0.9109 for its solution; lower is better on this measure. The authors cautioned that underlying models may already know public optimization techniques, so the experiments do not establish wholly independent rediscovery.⁠[6]

Across 235 GPU-kernel tasks, Recursive’s mean score rose from 0.699 to 0.754. Some candidate solutions exploited cached outputs, persistent state or timing weaknesses; added checks reduced these problems without guaranteeing that all errors had been found. The NanoGPT result still awaited testing on the official Prime Intellect hardware at publication, rather than being an officially confirmed leaderboard result.⁠[6]

The companion GitHub repository publishes training scripts and selected kernel implementations under Apache 2.0, retaining notices for upstream MIT-licensed code. Only ten of the 235 kernels were released to preserve future evaluation integrity. These are experimental artifacts, rather than the complete automated-research system.⁠[8]

On July 28, 2026, Recursive announced a $410 million multi-year agreement with Amazon Web Services to run its automated AI research system on AWS. The companies plan to support many parallel research loops and jointly develop infrastructure for large-scale automated research. The announced amount covers a multi-year service and development arrangement, rather than a confirmed one-time payment.⁠[9]

In a Latent Space interview published on September 14, Socher explained why he started Recursive separately: You.com had focused on search, search APIs and web answers, while he wanted to return to frontier model research and automate researchers’ idea-generation, implementation and validation work. He confirmed eight co-founders and identified Josh Tobin as CTO.⁠[10]

Socher described AI research itself as the initial target, including more efficient training, greater automation and improved inference, rather than starting with the physical sciences. These were statements of research priorities and future ambitions, not evidence that the company had achieved general superintelligence or automated discovery across scientific fields.⁠[10]

He said an early system was working but still required greater capability and scale to reach his intended level of recursive self-improvement, distinguishing the company’s ambition from completed experimental results.⁠[11]

In October 2026, Recursive introduced Flint for coding agents running on local hardware. The company described a development effort lasting under four months, combining a small human team with automated research across data, training, infrastructure and inference optimization. This account describes the company’s own development process.⁠[12]

The model card lists 39.8 billion total parameters, approximately 2.5 billion active per token, text input and output, reasoning and multi-turn tool use. Its documented context configuration is 262,144 tokens. The Hugging Face repository provides BF16 weights, a tokenizer and configuration; quantized variants are separate. The weights carry the Apache 2.0 license.⁠[13]

The published coding evaluations use specified execution conditions, and setups differ across developers. Local operation also depends on hardware, quantization and the inference engine. The release results do not establish unrestricted self-improvement by the full research system.⁠[13]