AI could accelerate science and uncover patterns too complex for humans to understand, economist Noah Smith predicts

AI could accelerate science and uncover patterns too complex for humans to understand, economist Noah Smith predicts

In his October 29 essay, Smith links future discoveries to the ability to absorb the scientific literature quickly and run many AI programs in parallel. To uncover complex relationships in nature, he proposes feeding models vast amounts of physical measurement data.

Science journalist John Horgan published his book The End of Science back in 1996. In a 2020 paper, economist Nicholas Bloom and colleagues documented declining returns on research. In chip manufacturing, crop yields, and efforts to reduce mortality from several diseases, improvements per unit of effort diminished over time.

Scientists spend years training, then work through the literature in their specialty. In Smith’s scenario, an AI program, once trained, could be run in many copies by adding chips and data centers. These programs, called AI agents, could pursue different projects simultaneously: reviewing the literature, comparing findings, and proposing hypotheses.

The scientific literature would help agents identify new questions, build on previous work, and borrow methods from other disciplines. Rapid analysis of experimental results would let them propose an explanation and choose the next experiment sooner. In a test by Prime Intellect, agents adjusted AI training settings, checked whether they could reach a specified level of performance in fewer steps, and used the results to choose their next attempt. Smith also describes a feedback loop: if discoveries increase the output of goods and services, some of the additional income could fund more computing capacity.

In an analysis cited by Smith, writer and futurist Ramez Naam distinguishes the ability to run more searches from the ability to solve a harder problem:

“A million independent searches can run simultaneously.”

When the choice of the next experiment depends on the result of the previous one, the experiments must proceed in sequence. Even dividing such a task among programs may require a new idea. Smith acknowledges that the prolonged search for new explanations may remain a human task.

Smith sees a limit to human collaboration in the patterns each participant can recognize and explain to others. He suggests that there may be

“laws of the Universe that humans cannot understand, but AI can.”

In a July essay, he called these “cloud laws”: causal relationships too complex for humans to explain. If such relationships exist, a model might detect them in cellular activity or plasma motion, predict these processes, and control them.

According to Smith, this search would require new physical measurements. He proposes building large numbers of sensors to measure activity in brains, cells, and other complex systems. AI would learn directly from this stream, processing more data than humans can describe and organize.