A new report sheds light on the secretive race to build self-improving artificial intelligence

As AI companies warn about the dangers of self-improving AI, a new report offers a rare look at how they're working to make it happen.

For months, Anthropic and OpenAI have warned that artificial intelligence may be on the verge of improving itself.

This milestone, known as "Recursive Self-Improvement" or RSI, could mean new AI capabilities come faster than ever. But the rise of RSI may also be very dangerous for humanity, potentially marking a stage where developers lose control of AI systems entirely.

Even as AI companies warn about the potential dangers of RSI, they have been remarkably coy about how to track it. A new report released Wednesday by Gladstone AI seeks to change that. This is one of the most detailed illustrations of how RSI can work and how AI companies are progressing towards it. The report is based on what the authors say are conversations with half a dozen researchers at top AI companies.

The report, which was shared exclusively with Ambassador Arjan ahead of its release, emphasizes that AI companies are "actively working" to automate all parts of the AI development pipeline, including generating new ideas, running experiments to test those ideas and interpreting the results. "Assuming no U.S. intervention or slowdown, the RSI will kick in within the next 3 to 12 months and possibly sooner," the report said, according to multiple Gladstone insiders it spoke to.

"There's a fair amount of [disaster] preparation and bioshelter buying going on among these researchers," says Edward Harris, Gladstone's chief technology officer.

The goal of the report, Goldstone executives tell Ambassador Arjan, is to give lawmakers a better understanding of how RSI works and, in turn, better ability to regulate it.

"Our goal here was to define reverse self-development so that people could have a rational conversation about it, especially in Washington," says Gladstone CEO Jeremy Harris. (These two are brothers.) "All these bills have been proposed to ban hyperdata, which are really interesting attempts to deal with this problem. But what we're most interested in is: What is the process that needs to be regulated?"

This report provides a simple overview of how the AI research pipeline currently works at AI companies and how those companies are beginning to automate the process.

For the first stage of training an AI model, known as pre-training, the emphasis is on finding ways to extract more intelligence from a model with less and less computational power, known in the industry as "computation".

AI companies use a portfolio approach to test these ideas. Researchers assign a small amount of computation to each idea and use it to train a series of models. Models that perform reliably better are selected for another round of training using larger amounts of computation.

Companies don't have infinite computing, which limits the number of promising ideas they can test at larger scales. This means that even if AI companies can train an AI that can generate ideas that seem promising, it's critical how many bad ideas it generates at the same time. If companies need to spend a lot of time or counting AI-generated good ideas from bad ones, it may make sense to continue to trust human researchers with the right intuition.

Conversely, AI companies with access to more computing can test more ideas – potentially allowing them to reach RSI first. With more computing for experiments, companies may get ideas that the Gladstone report describes as the rarest and most desirable of AI companies: those that perform better because more computing is directed at them, even if they look perfect at smaller scales.

Jeremy Harris says the ability to recognize when these ideas are being found, and especially whether AI is helping to find them, is a critical part of the process that regulators currently have little visibility into.

One characteristic that makes RSI adjustment difficult is that it is a gradual process. "That's actually one of the reasons we went to great lengths to describe it in so much detail," says Edward Harris. So that one can point to a part of this flow and say: Man must be there.