Gladstone AI publishes report on labs’ push to make AI improve itself

AI companies are trying to automate not just writing code, but the work of discovering how to build better AI. A Gladstone AI report published October 14 describes that effort, according to TIME’s coverage. Based on conversations its authors say they held with six researchers inside leading labs, it also presents an aggressive forecast.

Automating the research loop

The process is called recursive self-improvement: AI helps improve AI, potentially allowing new capabilities to arrive faster. Gladstone’s report describes companies working to automate the research pipeline, rather than presenting a single finished system that has crossed that threshold.

The report identifies three parts of that pipeline that companies are actively working to automate:

1. Generating ideas for improving models.
2. Running experiments to test those ideas.
3. Interpreting the results to understand what worked.

Its account begins with pretraining, the initial phase of training a model. Researchers look for ways to get more intelligence from less computing power, often called compute. They test a portfolio of ideas, assigning each a small computing budget to train a series of models.

Ideas that produce models which reliably improve as training grows are selected for another round with more compute. That selection step connects a promising small experiment to a much more expensive test. Automating idea generation alone does not remove the need to decide which experiments deserve those larger budgets.

More ideas can mean more filtering

Computing capacity limits how many promising ideas a company can test at larger scales. TIME’s explanation highlights a less obvious constraint: the bad ideas generated alongside the good ones. If sorting them consumes too much time or computing power, relying on human researchers with strong intuition may still make more sense.

More compute could therefore give a company an advantage in reaching self-improvement, by letting it test more possibilities. It could also help uncover ideas that look unsuccessful in small experiments but improve when given much larger computing budgets. Gladstone describes those as especially rare and desirable.

Our goal here was to define recursive self-improvement well enough that people can have a sane conversation about it, especially in Washington

Jeremie Harris, Gladstone AI CEO, speaking to TIME

A forecast—and a proposed place for oversight

Most researchers Gladstone consulted believe recursive self-improvement will begin within three to 12 months, possibly sooner, according to the report. That expectation assumes no U.S. government intervention or slowdown. It is an attributed forecast from a small group of insiders, not confirmation that the feedback loop has begun.

The stakes extend beyond faster research. TIME describes the potential danger as a point at which developers could lose control of AI systems. Gladstone’s executives say their purpose is to help lawmakers understand the process well enough to regulate it, rather than debate only the eventual arrival of superintelligence.

Jeremie Harris says lawmakers currently have essentially no visibility into whether companies are finding the most valuable ideas—and whether AI is helping find them. That makes the selection and interpretation of experiments relevant to oversight, not just the amount of automated work being performed.

Gladstone chief technology officer Edouard Harris describes self-improvement as gradual, which complicates regulation. The report’s detailed breakdown is intended to let policymakers identify particular steps where a person must remain involved. Its regulatory focus is therefore on where humans stay in the research process, not only on a future system’s capabilities.