| Wade into the conversation around artificial intelligence, and the one thing you won't find is consensus. Large language models like ChatGPT have mastered advanced reasoning, or they're mindless parrots. AI progress is accelerating, or it's running up against fundamental limitations. AI will soon usher in a golden age of scientific discovery, or it will undermine the very foundations of science. Why can't researchers agree on where AI technology is headed — or even where it currently stands?
One reason for the divergent views is that AI is unlike other technologies in one key respect. Engineers usually build machines by specifying exactly how they work. But AI systems don't receive this kind of explicit instruction. Instead, they learn by trial and error, with the help of loads of data. Through this process, they evolve internal mechanisms that enable them to accomplish a stunning variety of different tasks, but nobody really understands how these mechanisms work. No wonder observers can't agree on what they're looking at.
But even though AI is strikingly different from most new technology, it didn't come from nowhere. Rather, it has roots in fundamental science that burrow back decades. And many of the scientific disciplines that were so important to the development of AI are now confronting the possibility that AI will radically transform the way research is done.
Last week, Quanta published a special issue on science in the age of AI. It's a truly interdisciplinary effort, featuring contributions from six writers with years of experience covering computer science, math, physics and biology — not to mention editors, illustrators, graphic designers and producers who made it possible. The special issue examines the roots of AI in the sciences, how researchers are trying to make sense of the technology today, and how they're grappling with what the future might bring.
What's in the Series The first part of the series explores the origins of AI. It's hard to understand the discourse around AI without a solid grasp of the terminology. Our glossary of the most important terms in AI covers everything from the basics of neural networks — the mathematical structures at the heart of state-of-the-art AI systems — to controversial terms like "reasoning" and "hallucinations." Of course, the lingo of AI research is a product of the history of the field: The neural networks used in AI systems, for instance, were originally inspired by the structure of the human brain. But in the decades since they were first invented, artificial neurons have diverged from their biological counterparts. Yasemin Saplakoglu's feature story dives into this history, examines how biological and artificial networks differ, and explores why some researchers are once again taking inspiration from neuroscience. But that isn't the only field that has inspired the design of neural networks: The physics of strange materials led to innovations in neural network design in the 1980s. Elise Cutts tells the story of that breakthrough, which was recently honored with the Nobel Prize in Physics.
The second part of the series turns from the history of AI to the present. Today, neural networks are the engines behind state-of-the-art image recognition systems. I worked with Quanta's graphics editor Mark Belan on a visual explainer that breaks down the basic mathematics behind how they learn. In another explainer, I survey the techniques researchers are now using to study the inner workings of large language models like ChatGPT, and the counterintuitive phenomena they've discovered. And a gripping oral history compiled by the writer John Pavlus dives into the seismic impact that the advent of ChatGPT had on natural language processing, the field at the intersection of linguistics and computer science.
Natural language processing may have been the first scientific discipline to be dramatically reshaped by AI. It likely won't be the last. The final part of our series explores how AI might soon change math and science. Scientists have used AI-based tools to process and analyze data for many years, but some are now starting to use it in a more sacrosanct part of the scientific process — deciding what questions to ask in the first place. Unsurprisingly, that's proved controversial. Gregory Barber reports on researchers' attitudes toward this new approach to science, ranging from enthusiasm to scathing critique. And Jordana Cepelewicz explores how AI advances have forced mathematicians to reckon with the fundamental question of what math is really about — and whether it might soon look more like experimental science or even literary criticism. Throughout our reporting, we asked researchers to weigh in on where AI progress is headed and what it means for science. To close out the series, we're presenting their responses in their own words. If there's any unifying theme, it's that the future of AI remains profoundly uncertain. | 
