Former OpenAI chief research officer Bob McGrew says automating AI researchers won’t, by itself, trigger runaway self-improvement. Even automated researchers still need computing hardware. Source
Speaking with Sarah Guo and Elad Gil on a No Priors episode uploaded November 19, 2026, McGrew called that hard-takeoff argument “straightforwardly wrong.” Now Arda’s CEO and co-founder, he argues that making intelligence cheaper leaves plenty of expensive physical work unfinished. Source
Faster researchers still need experiments
Recursive self-improvement describes a proposed feedback loop: AI helps develop more capable AI, which can then accelerate the next round of development. McGrew challenges the assumption that eliminating researcher labor keeps multiplying the speed of discovery. Researchers have to test their ideas, and those tests require physical resources. Source
Recalling OpenAI in 2024, he said the constraint wasn’t how many researchers the company could hire. “It was the amount of GPUs that we had on hand.” The location of that computing capacity and the networks connecting it mattered, too, he said. Source
He expects substantial gains from research automation: better experiment selection, engineering and scheduling could extract more discoveries from available hardware. But he describes an S-curve, a large gain that eventually levels off as the system approaches its limits. Source
“It's a big one-time benefit, but it doesn't kick in a self-improvement cycle,” McGrew said. Source
In biology, he argues, cheaper hypotheses put more pressure on laboratories’ capacity to test them. His construction thought experiment starts with a need for more bulldozers to build data centers. An AI could propose a bulldozer factory—but clearing the site requires bulldozers. Congratulations: the bottleneck now has a factory blueprint. Source
Gil asked whether sufficiently good simulations could eliminate some physical experiments or reduce how many are needed. McGrew argued that even if simulations cut the number of required experiments tenfold, faster digital work would eventually run up against physical testing capacity again. Source
Arda’s 90-minute example is planning, not production
McGrew’s proposed response is also his business. Arda is building software and AI tools for manufacturing engineers, with the goal of letting users upload a product design and ultimately reconfigure an automated production line in seven days, he said. Source
At defense manufacturer Anduril, McGrew said, an engineer used Arda’s software to complete an actuator’s virtual build, a manufacturing-planning task, in 90 minutes, compared with the few days it normally took. Source
McGrew described the first Anduril user’s experience as “a tremendously frustrating process.” Arda’s product lead sat beside the engineer, observed problems and fed improvements back into the software, he said. The 90-minute figure is McGrew’s account of one engineer’s task, not an independently verified benchmark. Source
McGrew also expects AI to help expand physical capacity through increasingly automated factory and machine-tool design. When Gil contrasted a digital singularity with a physical one, McGrew replied: “not even versus,” but “just alongside.” Source
1 comment
that bulldozer thought experiment is both hilarious and apt