KorvaJoin the waitlist
Tech

AI Doing AI Research Could Shrink Years of Progress Into Months

Automating coding, debugging, and training could create a self-reinforcing cycle that starts virtual and then spills into the real world.

AI image
Dwarkesh Podcast · Paul Christiano – We could be 12 months away from AIs that can overthrow the government

The fastest engine for smarter AI may be AI itself, put to work as a researcher.

Paul Christiano, an AI safety researcher who leads a group called ARC, argues that automating AI research could compress what looked like four to eight years of progress into months. On the current trend alone, he estimates it might take about four years to build systems that reliably outcompete people, with more confidence by eight.

Automation changes the math by adding both quality and quantity of labor. That means AI systems writing code, debugging machine learning runs, and doing the other hands-on work that drives AI development.

The loop starts where AI is already strongest, inside the data center. Those are virtual tasks with clear right answers, such as improving software, designing better learning methods, and building new training environments to test hard skills.

Christiano says acceleration may already be around two times, and he describes faster growth as a scenario he considers fairly likely, not a finished result. In that scenario, he expects the pace of progress could double in eight months, then in four months, then in two months, then in one month, for an overall speedup of at least ten times.

Christiano expects that, for a while, developers chasing speed would care less about how good an AI lawyer is and more about how good it is at producing better architectures, better optimizers, and better ways to learn. In that stretch, builders would likely not care about most real world jobs, except for work that helps construct useful training environments.

Once that narrow loop is spinning fast enough, spreading the gains gets cheap. The same research engine could then be turned toward messy, real world tasks, which is why years of expected progress could arrive almost at once.

Questions people ask

What does AI research work actually involve?

It is mostly software and experimentation. Researchers write code, fix failed training runs, design learning algorithms, and build simulated tests that measure whether a new system is actually improving.

Why would this acceleration start with virtual tasks?

Virtual tasks have clear scores, so progress is easy to verify. A program either runs, fixes a bug, or earns a higher test score, while real world work like persuading people or running a lab is messier to judge.

Would AI researchers stop caring about other skills?

Christiano argues that builders focused on speed would first optimize the research loop itself. Broader abilities, from law to logistics, would come later, once the system can quickly convert its research gains into real world skills.

More in Tech