Acemoglu expects productivity gains from generative AI to be relatively modest, because developers at big tech companies have focused on using AI to replace humans through automation and “online monetization” in search and social media. To have a bigger impact on productivity, he argues, AI needs to be useful to a broader segment of the workforce and relevant to more parts of the economy. The key is that AI needs to be used to create new types of jobs, not just replace workers.
Acemoglu argues that generative AI can be used to augment workers' capabilities, for example, by providing real-time data and reliable information for many types of jobs — think intelligent AI agents familiar with the intricacies of factory floor production. But he writes that “these benefits will remain elusive unless there is a fundamental reorientation of the (technology) industry, including significant changes to the architecture of perhaps the most common generative AI models.”
It's tempting to think that today's large foundational models can simply be tweaked with the right data to make them more widely usable across industries, but in reality, we need to rethink our models and how they can be deployed more effectively in a broader range of applications.
Creating progress
Take manufacturing, for example. For many years, manufacturing has been one of the key sources of productivity growth in the US economy. Even today, it accounts for a large portion of US research and development spending. And while the recent increase in automation and use of industrial robots might suggest that manufacturing productivity is improving, this is not the case. For somewhat mysterious reasons, US manufacturing productivity has been in a dire state since about 2005, which has played a disproportionate role in the overall productivity decline.
The promise of generative AI to revive productivity is that it will help integrate everything from early material and design choices to real-time data from sensors embedded in production equipment. Multimodal capabilities could allow a factory worker, for example, to take a photo of a problem and ask an AI model for a solution based on that image, the company's operating manual, relevant regulatory guidelines, and reams of real-time data from the machinery.
At least, that's the vision.
In reality, efforts to bring today's foundational models to design and manufacturing are still in their early stages. To date, AI has been used in “narrow areas,” says Faez Ahmed, a mechanical engineer at MIT who specializes in machine learning, such as scheduling maintenance based on data from specific equipment. In contrast, generative AI models could, in theory, be useful for everything from using real-world data to improve initial designs, to monitoring steps in manufacturing processes, to analyzing performance data on the factory floor.
In a paper published in March, a team of MIT economists and mechanical engineers (including Acemoglu and Ahmed) identified many opportunities for generative AI in design and manufacturing, then concluded that “current (generative AI) solutions cannot achieve these goals due to several key deficiencies.” Chief among the shortcomings of ChatGPT and other AI models is their inability to provide reliable information, their lack of “relevant domain knowledge,” and their “lack of awareness of industry standard requirements.” These models are also not designed to handle the spatial challenges of manufacturing floors or the different types of data created by production equipment, including old machines.
The biggest problem, Ahmed says, is that existing generative AI models lack the right data. These models are trained on data collected from the internet, which “knows a lot more about cats and dogs and multimedia content than it does about how to actually operate a lathe,” he says. “The reason these models perform relatively poorly on manufacturing tasks is because they've never seen manufacturing tasks.”
Accessing this data is difficult, because much of it is proprietary. “Some people have a real fear that the models will take their data and run with it,” he says. A related issue is that manufacturing requires precision and adherence to strict industry and government guidelines. “If the system isn't accurate and you can't trust it, people aren't going to use it,” he says. “And it's a chicken-and-egg problem: the models aren't accurate because we don't have the data.”
