Companies like Anthropic and OpenAI released new AI systems late last year that were particularly good at writing computer code.
In recent months, the technology has rapidly reshaped the way Silicon Valley engineers build, test and modify new software applications. If an AI system can write code, it can help speed up the development of things as diverse as word processors and social media applications.
Now, many of the world’s leading researchers believe that AI will soon be powerful enough to improve itself with little or no help from human developers.
“AI is code. And now AI can code,” said veteran researcher Richard Socher. “The folders are there.
Dr. Socher recently founded a company with seven other researchers to pursue this mind-bending goal, often called “recursive self-improvement.”
His start-up, Recursive Superintelligence, has raised more than $650 million from venture capital firms, including GV (formerly Google Ventures), Greycroft and chip giants Nvidia and AMD. The six-month-old company, which has offices in San Francisco and London, has fewer than 30 employees. But now it is valued at more than 4 billion dollars.
The company should not be confused with Recursive Intelligence, which pursues a similar goal and is also valued at $4 billion. Prominent AI startups Anthropic and OpenAI are also pursuing recursive self-improvement, a decade-long obsession among Silicon Valley technologists.
Dr. Socher, who is also the CEO of artificial intelligence start-up You.com, was previously head of artificial intelligence research at business software maker Salesforce. Its seven co-founders include prominent researchers from many leading AI companies, including Josh Tobin, Jeff Clune, and Tim Shi, all of OpenAI, and Yuandong Tian of Meta.
Many of these researchers specialize in a type of AI development called “openness”. This involves building software systems that can run for days, months, or even years while tracking goals set by researchers.
Recursive superintelligence also hired Peter Norvig, who spent 25 years as director of research at Google and co-authored the textbook on artificial intelligence (“Artificial Intelligence: A Modern Approach”), which has been a university standard for three decades.
Recursion, a term common among mathematicians and computer programmers, refers to a mathematical function that feeds on itself. After a recursive procedure generates information, it uses that information to generate something else—and so on.
Although many researchers are optimistic about the idea of recursively improving AI, others point out that current technology is far from the point where humans can be removed from the loop. People like Dr. Socher – must keep creating new ideas that propel the development of artificial intelligence forward.
But the goal is to push more and more work onto machines, including generating new ideas.
OpenAI said it is now building an “automated AI researcher”. By fall, the company hopes to have a system that can handle the work of a “less experienced” researcher, said Sam Altman, CEO of OpenAI. Similar efforts are underway at other leading companies.
Dr. Socher said his startup would take years to build the kind of technology he and his co-founders envisioned. The company hopes to eventually use the technology in other areas, such as drug discovery and other types of biological research.
(The New York Times has sued OpenAI and Microsoft, alleging copyright infringement of news content related to AI systems. Both companies have denied the suit’s claims.)
