Anna Goldie and Azalia Mirhoseini, the founders of Ricursive Intelligence, look at data at the start-up’s offices in Palo Alto, California. Even as many financial analysts and industry insiders warn of an artificial intelligence bubble, vast amounts of money continue to pour into the field. — Cayce Clifford/The New York Times
SAN FRANCISCO: One new startup is called Recursive with an “e”. Another is called Ricursive with an “i”. They are trying to do the same thing: Build artificial intelligence that can improve itself without the help of humans, an obsession of Silicon Valley technologists for decades.
Ricursive Intelligence, based in Palo Alto, California, is working with the specialized computer chips that power today’s chatbots. Founded by two former Google researchers, Anna Goldie and Azalia Mirhoseini, Ricursive aims to build AI systems that can improve the design of these enormously complex chips.
If AI systems can produce better chips, they argue, the chips will produce better AI systems. And then the process would repeat on and on as technology got better and better.
“The idea of a recursive self-improvement loop is what inspires us,” said Goldie, who did similar work with Mirhoseini at Google.
Ricursive has raised US$335mil from venture capital firms like Sequoia, Radical Ventures, Lightspeed and DST Global. Although it is less than a year old and has fewer than 10 employees, it is valued at US$4bil.
The company is among several new AI startups that have raised enormous amounts of money in recent months. Just last week, Humans&, founded in San Francisco by former researchers at labs like Anthropic and Elon Musk’s xAI, raised US$480mil. It is just three months old and valued at US$4.48bil.
Even as many financial analysts and industry insiders warn of an AI bubble, vast amounts of money continue to pour into the field. This is partly because the raw computing power needed to build AI technologies is so expensive. If investors want to bet on a new idea, hundreds of millions of dollars are increasingly the table stakes to get into the AI game.
Recursion is a term commonly used by mathematicians and computer programmers. It refers to a mathematical function or procedure that feeds itself. After a procedure generates some information, it uses that information to generate something else. That process can go on forever.
That mathematical idea has inspired AI researchers for decades. Rather than building just a mathematical function that feeds itself, they aim to build an AI system that feeds itself.
In 2017, as the latest wave of AI development first gained steam, Google built technology called AutoML. ML was short for “machine learning,” referring to computer algorithms that learn skills by analyzing data. With AutoML, Google took this idea a step further. It built a machine-learning algorithm that learned to build other machine-learning algorithms.
At OpenAI, the maker of ChatGPT, researchers are building what they call an “automated AI researcher.” By the fall, they hope to have a system that can do the work of a less experienced researcher, before steadily improving on the technology, the company’s CEO, Sam Altman, said.
That is similar to the aim of another new startup, Recursive AI, founded by Richard Socher, who oversaw AI research at the cloud computing giant Salesforce. His startup has yet to publicly announce itself, but its mission is already a topic of discussion across Silicon Valley’s tight-knit community of AI researchers. Recursive AI is also valued at US$4bil, according to a person familiar with its latest funding round who spoke on the condition of anonymity. The news was first reported by Bloomberg.
Although technologies as far back as Google’s AutoML have shown that AI can help improve AI, these efforts are still a very long way from a future where humans can be removed from the process, said Div Garg, CEO of AGI, a San Francisco startup that is working to build increasingly intelligent computer technologies.
“They work well for very specific tasks,” he said.
At Google, Goldie and Mirhoseini built AI technology that could improve the design of the company’s in-house computer chip. Called the tensor processing unit, or TPU, the chip was designed for building and running AI technologies.
Now, Ricursive plans to help other companies hone their chips in similar ways. And as the years pass, its larger goal is to create a virtuous circle where the chips and the AI evolve alongside each other.
“The first phase of the company is just to accelerate chip design,” Goldie said. “But if we have the ability to design chips very quickly, why not just use that ourselves? Why not build our own chips? Why not train our own models? Why not coevolve them?” – ©2026 The New York Times Company
This article originally appeared in The New York Times.
