What happened?

An essay published in MIT Technology Review argues that AI's route to accelerating science differs from the common assumption. Its starting point is an observation: every few decades, someone announces that science has reached its end. In 1903 the physicist Albert Michelson wrote that the "facts of physical science have all been discovered"; in the 1980s Stephen Hawking predicted theoretical physics might be finished by the end of the century.

With the arrival of AI, that feeling is in the air again — this time accompanied by a Nobel Prize. In 2024, Demis Hassabis and John Jumper of Google DeepMind were awarded part of the Nobel in chemistry for AlphaFold, a neural network that predicts three-dimensional protein structures by learning from thousands of experimentally measured shapes.

The condition that made AlphaFold possible

This devilish problem had resisted systematic attacks for half a century; AlphaFold seemed to have solved it, and the world became fixated on the promise of its approach. Hassabis and his team called AlphaFold "the template for how AI can accelerate all of science to digital speed". A wave of startups building foundation models for biology, chemistry and materials discovery raised billions of dollars.

The essay's objection lands here: AlphaFold is a profound achievement, but the conditions that produced it are rare. The primary condition for its success was the Protein Data Bank, a data set of roughly 170,000 experimentally validated protein structures on which DeepMind's team could train its model.

The cost of the data

  • Data set: roughly 170,000 validated protein structures
  • Time to assemble: 53 years of international scientific cooperation
  • Estimated experimental cost: roughly $21 billion
  • Nobel connection: more than 25 Nobel Prizes have relied on protein crystallography, the key experimental technique

Efforts of that scale, the piece notes, are infamously difficult to fund, next to impossible to coordinate and hugely time-consuming to execute; they have often been unsuccessful as a result.

The barrier that is too little discussed

Even in fields with the requisite cohesion and resources, and where the relevant data are not rendered inaccessible by commercial ownership, another barrier goes under-discussed: the scientific impossibility of generating comparable data.

In the case of protein structures, the key experimental technique — protein crystallography — is an unusually replicable and dependable tool, so much so that over 25 Nobel Prizes have relied on it. Other fields have no measurement technique of comparable reliability.

Why does it matter?

The conclusion is that AI will bring extraordinary changes to science, but AlphaFold may not be the best template for that metamorphosis. Meeting those conditions in other fields will take time measured in decades, not years. Instead, the acceleration of science will come about thanks to another approach: AI agents.

In practice the distinction means this: a data-driven model waits for the data to exist. A reasoning agent can take part in the process of producing that data — deciding which experiment to run and which hypothesis is worth pursuing. The first depends on a historical accumulation; the second joins the work of building it.

What is not settled

This is an essay rather than a news report, and it offers no numerical result on how well agents perform in scientific discovery. Its case rests on an observation about the limits of the current approach and on the logic of the alternative.