Compute Advantage Drives AI's Self-Improving Race

Leading AI companies are trying to automate every stage of development. Researchers interviewed for a Gladstone AI report say recursive self-improvement could begin within 3 to 12 months, and companies with more computing power may test more ideas and get there first.

Who Controls the Pipeline

The report describes companies working to automate the process, from generating research ideas to running experiments and interpreting results. Most of the half dozen researchers inside top AI companies who spoke with the authors believe recursive self-improvement, or RSI, could start within 3 to 12 months. They said it could happen sooner if there’s no U.S. government intervention or slowdown.

During pretraining, companies search for ways to get more intelligence from a model while using less computing power, known in the industry as “compute.” Researchers test portfolios of ideas by assigning a small amount of compute to train a series of models. Ideas that reliably improve performance as models scale move into another training round, where teams use more compute.

Computing resources are limited. That limits how many promising ideas companies can test at larger scales. AI can generate ideas, but researchers still have to sort useful ones from poor ones. That filtering can take enough time or compute that companies may prefer human researchers with sound intuition, the report says.

Companies with more compute can test more ideas and may reach RSI first, the report says. More computing power may also help them spot ideas that look like failures at smaller scales but perform better when teams apply additional compute. The report describes a competitive advantage tied directly to access to computing resources; it doesn’t identify specific company profits or financial figures.

What the State Can See

Jeremie Harris said legislators currently have “essentially zero visibility into” how companies identify promising ideas and whether AI helps find them. The report’s executives said they want lawmakers to better understand the process and how it might be regulated.

Proposed bills to ban superintelligence are “really interesting attempts to grapple with this problem,” Jeremie Harris said. He also asked, “what is the process that needs to be regulated?” The report presents gradual RSI development as a challenge for regulation. Edouard Harris said it’s difficult to regulate because it develops gradually.

“This is actually one of the reasons why we went to pains to describe it in so much detail,” Edouard Harris said, “so that it is possible to point to a part of this process and say: a human needs to be there.” The proposals described in the report concern banning superintelligence, while the report’s executives are asking lawmakers to identify specific steps in the development process for regulation.

Risk Without Public Visibility

Anthropic and OpenAI have warned for months that AI may be nearing the point at which it can improve itself. RSI could bring new capabilities faster, but developers could also lose control of AI systems entirely, the article reports.

“There’s a fair amount of [disaster] prepping and bio-bunker purchasing going on” among researchers, Edouard Harris said. TIME received the report exclusively before publication on Wednesday. It describes how RSI might work and how companies are progressing toward it.

The article identifies researchers, company executives and lawmakers in the debate. It reports no union, worker organization or organized labor response, and gives no figures for company profits, worker pay or employment effects. Instead, its central account of the race tracks who can test more ideas: companies with greater access to compute.