Explore thousands of variations in hours instead of months
A new AI system called ERA, described in Nature, has demonstrated the ability to generate expert-level scientific software across disciplines ranging from epidemiology to neuroscience—sometimes surpassing the benchmarks set by human researchers. Built on a large language model guided by tree search algorithms, ERA treats code creation as an iterative optimization process, drawing on scientific literature to explore thousands of variations in hours rather than months. Its emergence marks a meaningful shift in the pace and accessibility of computational research, though it also invites the enduring question of whether accelerating the tools of science is the same as deepening our understanding of the world those tools describe.
- ERA outperformed the official CovidHub Ensemble on COVID-19 hospitalization forecasting, achieving a mean Weighted Interval Score of 26 against the benchmark's 29—a result that unsettles assumptions about where expert judgment ends and machine capability begins.
- The system generated 40 novel bioinformatics approaches for single-cell RNA sequencing, with one variant improving on published methods by 14 percent, compressing what might take a research team months into a matter of hours.
- By integrating knowledge from papers, textbooks, and search results, ERA can operate in domains where no public code exists—raising the stakes around who controls these tools and how they are deployed in sensitive fields.
- Experts caution that ERA optimizes predictions without reasoning about causes, meaning its outputs can be powerful without being understood—a distinction that matters enormously when wrong forecasts carry real-world consequences.
- The system is landing as both a democratizing force for under-resourced research teams and a source of unresolved safety questions, sitting at the edge of a broader reckoning about what it means to automate scientific work.
A new artificial intelligence system called ERA has demonstrated the ability to write scientific software at a level that rivals—and sometimes exceeds—human experts. Published in Nature, the system pairs a large language model with a tree search algorithm to automatically generate, test, and refine code across bioinformatics, epidemiology, geospatial analysis, neuroscience, and numerical computation. What sets ERA apart is its capacity to draw on scientific papers and textbooks, then systematically explore thousands of code variations to find the best-performing solution.
The problem ERA addresses is familiar to any research team: building specialized scientific software is slow, costly, and demands deep expertise. ERA reframes code generation as an optimization loop—producing candidate programs, evaluating them against performance metrics, and iteratively improving the most promising ones. It can modify existing software, recombine approaches across methods, and adapt to nearly any research task.
Tested across six scientific benchmarks, the results were striking. In bioinformatics, ERA produced 40 new approaches for analyzing single-cell RNA sequencing data, with one variant improving performance by 14 percent over published methods. In epidemiology, it generated 14 COVID-19 forecasting strategies that outperformed the CovidHub Ensemble—the model used by U.S. public health authorities—by combining statistical trend analysis with disease-spread modeling in ways neither approach could achieve alone.
For time-sensitive research or teams with limited resources, ERA's ability to compress months of work into hours carries real significance. Yet the study's authors are deliberate in marking its limits: optimizing a predictive model is not the same as understanding why something works. ERA finds patterns and builds forecasting systems; it does not reason about mechanisms or causality. And in sensitive domains, deploying powerful computational tools without deep expertise introduces risks that faster software alone cannot resolve. ERA is a genuine acceleration of certain kinds of scientific work—but not a substitute for the reasoning that gives that work meaning.
A new artificial intelligence system called ERA has learned to write scientific software at a level that rivals—and sometimes exceeds—the work of human experts. The system, described in a study published in Nature, combines a large language model with a tree search algorithm to automatically generate, test, and refine code across a range of research domains: bioinformatics, epidemiology, geospatial analysis, neuroscience, and numerical computation. What makes ERA distinctive is not just that it can write code, but that it can do so while integrating knowledge from scientific papers, textbooks, and search results, then systematically explore thousands of variations to find the best solution.
The challenge ERA was built to solve is real and familiar to any research team: writing specialized software for scientific problems is slow, expensive, and demands deep expertise. A researcher might spend weeks or months developing a single analysis pipeline. ERA collapses that timeline. The system treats code generation as an optimization problem. It creates multiple candidate programs, evaluates them against a performance metric, then rewrites and improves the most promising ones in a feedback loop. Unlike earlier automation tools that either generate code from scratch or work within rigid templates, ERA can modify existing software, recombine approaches from different methods, and adapt to almost any research task—from data preparation to complex simulations to advanced mathematics.
When researchers tested ERA across six scientific benchmarks, the results were striking. In bioinformatics, the system generated 40 new approaches for analyzing single-cell RNA sequencing data, with one variant improving performance by 14 percent over previously published methods while preserving the biological signals that matter most. In epidemiology, ERA produced 14 forecasting strategies for COVID-19 hospitalizations in the United States that outperformed the official CovidHub Ensemble, the benchmark model used by public health authorities. ERA achieved a mean Weighted Interval Score of 26 compared to the Ensemble's 29—lower scores indicate better performance. The system accomplished this by combining different modeling approaches: pairing statistical trend analysis with disease-spread models to capture patterns neither alone could see.
What's particularly notable is how ERA achieved these results. The system didn't rely on a single clever algorithm. Instead, it systematically explored variations, learning from external knowledge and recombining successful elements from different methods. When researchers instructed ERA to build its own boosted decision tree library without using existing software packages, it did so—and the results held up under manual verification. The system performed well even on tasks where no public code existed to learn from, suggesting it had developed genuine capability rather than simply remixing existing solutions.
The implications are significant for how research gets done. ERA can generate expert-level software in hours or days instead of weeks or months. For teams working on time-sensitive problems—like pandemic forecasting—that acceleration matters. For researchers in resource-limited settings, it could democratize access to sophisticated computational methods. But the authors of the study are careful to note what ERA is not. Optimizing a predictive model is not the same as discovering why something works. ERA excels at finding patterns in data and building systems that predict outcomes. It does not reason about mechanisms, causal relationships, or underlying theory. And there are safety concerns: if ERA makes it easier to deploy advanced computational models without deep expertise, that could create risks in sensitive domains where a wrong prediction carries real consequences. The system represents a genuine shift in how scientific software gets built, but it is a tool for accelerating certain kinds of work, not a replacement for the reasoning that drives scientific understanding.
Citações Notáveis
Optimizing empirical predictive models is not the same as full scientific discovery, which also requires reasoning about mechanisms, causal relationships, theories, and mathematical frameworks.— Study authors, Nature
A Conversa do Hearth Outra perspectiva sobre a história
So ERA writes code that outperforms human experts. Does that mean the experts are obsolete?
Not quite. ERA is very good at one specific thing: optimizing code to maximize a measurable score. It can explore thousands of variations faster than any human could. But it doesn't ask why something works, or whether the measurement itself is capturing what matters.
Give me a concrete example.
The COVID forecasting work. ERA combined statistical trend analysis with epidemiological models—two different approaches—and that hybrid worked better than either alone. A human expert might have thought of that combination, but it would take time and intuition. ERA just tried it systematically.
And the bioinformatics work, where it improved performance by 14 percent?
It generated 40 different approaches to the same problem. One of them was better. The key is that it preserved the biological signals—the actual science—while optimizing the technical performance. That's the constraint that mattered.
What's the catch?
ERA optimizes what you ask it to optimize. If you measure the wrong thing, ERA will get very good at the wrong thing very quickly. And it doesn't understand causation. It finds patterns. For real scientific discovery, you need both.
So it's a tool for researchers, not a replacement for them.
Exactly. It's a tool that compresses months of coding into days. That frees researchers to think about bigger questions. But someone still has to ask the right questions.