For all the obsessive nattering and nail-biting about artificial intelligence replacing lawyers, call-center workers, or copywriters—even policy analysts!—the bigger possible impact should be welcomed by all: cracking the frontiers of science. A new report from Epoch AI, a research group focused on forecasting advanced AI, argues that if current trends in AI scaling (industry jargon for steadily training ever-bigger models on more data, computing power, and energy) persist to 2030, the technology could reshape R&D in ways that make office-task automation look like a statistical and socioeconomic sideshow.
From the report: “We argue that AI scaling is likely to continue through 2030, despite requiring unprecedented infrastructure, and will deliver transformative capabilities across science and beyond.”
The numbers here are eye-popping. By 2030, the largest AI models could require 1,000 times today’s compute, training clusters costing hundreds of billions of dollars, and city-scale electrical demand. Individual training runs themselves are likely to cost in the billions. Such scale might sound science fictional, at least until you recall that even three years ago, today’s frontier systems looked implausible.
Here’s why such investments might pay off: Extrapolating from today’s benchmarks—standardized tests researchers use to measure AI progress on specific scientific and technical tasks—the report suggests that by 2030 AI could plausibly “implement complex scientific software from natural language, assist mathematicians formalizing proof sketches, and answer open-ended questions about biology protocols.”
In other words, warp-speed science (even if actual warp-drive engines take a bit longer). Accelerating progress in mathematics, molecular biology, and weather prediction is within sight, the report concludes. That’s because, in effect, many scientific disciplines may soon get their own version of superefficient coding assistants. Epoch predicts that AI assistants are “at minimum” likely to improve day-to-day productivity by 10–20 percent, at least within non-experimental work tasks. That finding may understate the upside since the figure comes from early Copilot trials in 2023–24, before more capable autonomous agents emerged.
Now this is the point AI worriers often miss: The potential economic upside dwarfs the usual discussion about automating white-collar drudgery. Productivity gains in scientific R&D wouldn’t just benefit researchers. They would eventually cascade through the entire economy, compounding into faster innovation-driven growth. (Keep in mind that most economic forecasts about the impact of AI don’t attempt to calculate the impact of faster scientific progress.)
There are caveats, including some with policy implications. Progress will be uneven across fields, possibly bottlenecked by data or power constraints. And as the report stresses, “deployment timelines are contingent on hard-to-predict sociotechnical choices.” Don’t expect blockbuster AI-invented drugs to hit pharmacy shelves by 2030. Under today’s sluggish approval pipelines, that’s too soon.
Still, if Epoch AI is directionally right and if government can do its part—hardly a done deal—the foundations will already be laid for an era where AI becomes a powerful engine of discovery. So instead of just asking how many jobs AI might take, it may be more useful to also ask what entirely new industries it could help create. The biggest economic return from AI advances will be enabling breakthroughs no human alone could achieve.
The post AI’s Greatest Trick May Be in the Lab, Not the Office appeared first on American Enterprise Institute – AEI.
