A new AI system developed by researchers from Google, Harvard and MIT offers a glimpse of something much bigger than another chatbot: machines that can search for discoveries themselves.

For the past several years, most of the conversation around artificial intelligence has centered on one question:

What jobs will AI replace?

Writers. Programmers. Accountants. Customer-service representatives. Analysts.

It may be the wrong question.

A far more consequential possibility is beginning to emerge from the world’s leading AI laboratories and universities.

Artificial intelligence may not simply automate the work humanity already knows how to do.

It may begin automating the process through which humanity discovers things it does not yet know.

A new study published in Nature provides one of the clearest demonstrations yet.

Researchers primarily from Google Research and Google DeepMind, working with scientists affiliated with Harvard University and MIT, developed an artificial-intelligence system called Empirical Research Assistance, or ERA. The project was co-led by Google researcher and Harvard professor Michael Brenner. (nature.com)

ERA is designed to tackle one of the less glamorous bottlenecks of modern science: experimentation.

Scientific breakthroughs rarely appear in a single flash of inspiration.

Researchers propose an idea. They build a model. They write software. They run an experiment. They examine the results. Something fails. They alter the model. They run it again.

Repeat.

Hundreds of times.

Sometimes for years.

ERA attempts to automate parts of that cycle.

An AI That Experiments

The basic idea is surprisingly elegant.

Give ERA a problem whose outcome can be measured.

The AI generates software intended to solve the problem.

The software runs.

Its performance receives a score.

ERA examines what happened, modifies the program, tries another approach and scores that one too.

Instead of following a single path, ERA uses a form of tree search that allows it to explore different branches of possible solutions. Promising ideas can be developed further. Dead ends can be abandoned. Earlier approaches can be revisited and recombined.

The system essentially conducts a massive computerized process of trial and error. (Nature)

There is an important difference between this and simply asking ChatGPT or another large language model to write some code.

ERA creates a feedback loop.

Generate. Test. Measure. Improve. Repeat.

And it can repeat the cycle far more relentlessly than a human researcher.

That matters because many scientific problems have enormous search spaces.

There may be thousands—or millions—of plausible ways to attack a problem.

A human scientist must choose which handful seem worth pursuing.

An AI researcher can potentially explore far more of them.

Then Something Interesting Happened

The researchers tested ERA across several fields, including genomics, epidemiology, neuroscience, geospatial analysis, time-series forecasting and numerical mathematics.

The results were remarkable.

In a benchmark involving single-cell RNA sequencing, ERA produced 40 methods that outperformed the best previously evaluated human-developed method.

In another experiment, the system developed 14 approaches that outperformed the CDC ensemble and all individual models included in the researchers’ comparison for forecasting COVID-19 hospitalizations. (Nature)

ERA also reached expert-level performance on problems involving satellite imagery, neural activity in zebrafish brains and difficult numerical calculations. (Nature)

Perhaps most importantly, researchers reported that the system could compress certain cycles of experimentation that might normally require weeks or months into hours or days. (Nature)

Think about what happens if that capability keeps improving.

Science Has Always Been Limited by Human Bandwidth

There are only so many scientists.

There are only so many hours in a day.

And there are vastly more hypotheses worth testing than humanity has the time or money to test.

That creates an invisible constraint on technological progress.

It isn’t necessarily that humanity lacks ideas.

Sometimes we lack enough experimentation.

Imagine a cancer researcher with 500 plausible hypotheses but enough time and funding to rigorously investigate five.

The remaining 495 largely disappear.

Now imagine an AI system capable of helping investigate hundreds of them.

The economics of discovery begin to change.

And once the cost of experimentation falls, previously uneconomic research becomes possible.

This could matter enormously for drug development, materials science, energy, climate modeling, biotechnology, engineering and nearly every computational science.

AI Is Beginning to Enter a Feedback Loop

ERA is not developing in isolation.

Google DeepMind’s AlphaEvolve uses large language models and automated evaluation to iteratively develop improved algorithms.

Google has reported that AlphaEvolve has helped optimize its data centers, improve aspects of chip design and AI training, discover improved matrix-multiplication algorithms and make progress on mathematical problems. (Google DeepMind)

By 2026, Google reported applications extending into quantum computing, DNA sequencing, power-grid optimization and other areas of scientific research. (Google DeepMind)

A pattern is emerging.

Large language models generate possibilities.

Computers test them.

Successful ideas survive.

Unsuccessful ideas disappear.

The system generates another generation of possibilities.

It resembles evolution occurring inside a computer.

And this leads to one of the most consequential potential feedback loops of the coming century:

Better AI creates better science.

Better science creates better chips, medicines, materials, batteries and algorithms.

Those technologies create better computers.

Better computers create better AI.

And the cycle begins again.

Humanity has experienced technological feedback loops before.

But we have never possessed an intellectual technology capable of participating directly in the invention process at massive scale.

This Does Not Mean AI Has Become a Scientist

There is an important reason for caution.

The ERA researchers themselves emphasize that optimizing a measurable result is not equivalent to genuine scientific discovery.

Science requires more than winning benchmarks.

Scientists attempt to understand why something happens.

They build theories.

They distinguish correlation from causation.

They design experiments capable of disproving their own ideas.

ERA is currently best suited to what its creators call scorable tasks—problems where a computer can objectively evaluate whether one solution is better than another. (Nature)

Reality is often messier.

You cannot place every scientific question onto a leaderboard.

But that limitation may also tell us where AI-driven discovery will advance fastest.

Anywhere nature gives us a measurable feedback signal becomes a candidate for accelerated machine experimentation.

Protein structures.

Molecular properties.

Energy efficiency.

Chip performance.

Disease forecasting.

Materials strength.

Battery chemistry.

Manufacturing processes.

Financial models.

Logistics.

Robotics.

Those areas contain enormous numbers of problems where proposed solutions can be generated, tested and scored.

The Coming Age of Machine-Speed Science

For decades we have measured technological progress largely by computing power.

Processors became faster.

Storage became cheaper.

Networks became larger.

Artificial intelligence introduces another variable:

the speed of experimentation itself.

Suppose AI eventually allows scientists to perform 100 times as many useful computational experiments.

Then AI hasn’t merely increased productivity.

It has expanded the effective population of researchers.

Push the idea further.

Imagine thousands of AI research agents operating continuously.

One reads newly published research.

Another generates hypotheses from it.

Another writes simulations.

Another analyzes experimental data.

Another searches for contradictions.

Another proposes an improved model.

Humans provide objectives, judgment, physical experiments and ultimately meaning.

Machines provide enormous intellectual search capacity.

The scientist of the future may therefore look less like someone manually conducting every calculation and more like the director of an enormous digital laboratory populated by artificial researchers.

What This Could Mean for Charlotte

The implications aren’t confined to Silicon Valley, Cambridge or university laboratories.

Charlotte sits at the intersection of industries likely to become increasingly computational: banking, healthcare, energy, logistics, construction and advanced manufacturing.

Companies that learn how to turn their problems into measurable optimization problems could effectively deploy AI researchers against them.

A hospital could search for better operational models.

An energy company could continuously optimize parts of its grid.

A bank could test new approaches to fraud detection and risk.

A manufacturer could explore thousands of engineering configurations before constructing a physical prototype.

The divide of the coming decade may therefore be less about which companies “use AI.”

Nearly everyone will.

The meaningful divide may be between organizations that use AI to perform existing tasks and organizations that use AI to discover better ways of doing things.

The second category could become dramatically more powerful.

The Real AI Revolution

We tend to imagine technological revolutions through visible products.

The automobile.

The television.

The smartphone.

The humanoid robot.

But the most important consequence of artificial intelligence may ultimately be much less visible.

It may happen inside laboratories and data centers, where millions of digital experiments quietly run while we sleep.

Some fail.

Some improve slightly.

Occasionally one discovers something no human researcher happened to try.

Then that discovery becomes the starting point for another search.

And another.

And another.

The researchers behind ERA conclude that scientific fields where solutions can be evaluated automatically may be approaching a substantial acceleration in progress. (Nature)

That possibility deserves far more attention than another AI chatbot.

Because once machines can participate meaningfully in discovery, technological progress itself becomes partially automatable.

The defining question of the AI age may therefore not be:

What can artificial intelligence do?

It may be:

What can artificial intelligence discover next?