Recursive Superintelligence raises $650 million to build self-improving AI models

Recursive Superintelligence Inc., a startup aiming to develop self-improving artificial intelligence models, was launched today with $650 million in funding.

Alphabet Inc.’s GV Fund and Greycroft led the round. The venture capital arms of Nvidia Corp. and Advanced Micro Devices Inc. participated. The investment values ​​the company at $4.65 billion, according to Recursive.

The company was founded earlier this year by former Salesforce Inc. chief scientist Richard Socher. He previously launched You.com Inc., a provider of application programming interfaces used by AI models to conduct online research. The startup was valued at $1.5 billion last year.

According to the New York Times, Recursive’s initial team consisted of Socher and six other staffers. The company currently has more than 25 employees in San Francisco and London. They are working to build so-called recursive self-improving superintelligence, or AI models that can discover new knowledge just like human scientists.

Current neural networks cannot perform basic research in a fully autonomous manner. As a result, Recursive’s top priority is building AI models that can improve its own code base. The company hopes that such models will help it discover ways to develop AI that is as effective as humans at scientific tasks.

The company’s AI will explore ways to improve itself by running simulations in an “open-ended process of automated scientific discovery.” According to Recursive, the model develops experimental ideas and tests them to validate the results. The company plans to develop guardrails to prevent the software from producing dangerous output.

Recursive says the experiments run by its AI models will focus on improving the harness as well as the code. Harness is a set of auxiliary programs that AI providers use to enhance the output of their algorithms. Additionally, Recursive’s system will explore ways to improve its training and inference infrastructure.

OpenAI Group PBC is already using the recently released GPT-5.5 model for this purpose. The company splits each inference request into so-called chunks and spreads those chunks across multiple graphics card cores to speed up processing. Until recently, the number of chunks included in a workflow was fixed. According to OpenAI, GPT-5.5 has developed more efficient parallelization techniques that increase token generation speed by more than 20%.

Some companies are using AI to power not only their inference workflows but also the underlying hardware. For example, the reflexive investor’s alphabet is design Its TPU accelerator works with the help of a neural network trained on the chip’s blueprint. Creator of recent systems launched A startup called Ricursive Intelligence Inc. aims to make similar technology available to other companies.

Recursive has not disclosed the machine learning techniques used to power its self-improving AI. Rival Ineffable Intelligence Ltd. also aims to develop models that can discover new knowledge. using Reinforcement learning. This is an AI modeling technique commonly used in large language modeling projects.

“We will start with AI research itself, but eventually we want to expand into physics, chemistry, and especially preclinical biology,” Socher said. I wrote In a post on

Image: Unsplash