What's Happening?
New research from MIT Sloan indicates that the scale of a firm's operations is a critical factor in making AI automation efforts financially viable. Unlike previous technologies where the marginal cost of production using AI was the primary focus, this
study highlights the substantial fixed costs associated with training AI models. These initial training costs do not increase with the scale of production, meaning that larger enterprises are significantly more likely to be early adopters of AI automation than smaller companies with fewer resources. The research, conducted by Danial Lashkari, Wensu Li, Neil Thompson, and Christina Qiu, specifically examined tasks automatable by computer vision, which account for approximately 3.5% of total economic tasks. They found that scaling AI accounted for 63% of its comparative advantage across computer vision tasks in 2023, a figure projected to remain around 60% by 2050.
Why It's Important?
This finding has profound implications for the competitive landscape of U.S. industries. The high fixed costs of AI model training create a barrier to entry for smaller businesses, potentially exacerbating the divide between large corporations and small-to-medium enterprises (SMEs). Larger companies, with their ability to absorb these initial investments and leverage AI across extensive operations, can achieve significant efficiency gains and cost reductions, further solidifying their market positions. This could lead to increased market concentration and reduced competition in various sectors. Conversely, SMEs might struggle to keep pace, potentially facing disadvantages in productivity and innovation. The research suggests that the economic benefits of AI are not evenly distributed, favoring those with the capacity for large-scale implementation.
What's Next?
The adoption rate of computer vision automation is projected to rise, with an unweighted adoption rate nearing 23% by 2035, and an employment-weighted rate approaching 60%. This increase is anticipated as computing prices fall and models become easier to fine-tune. However, the researchers predict that real output gains from computer vision automation will eventually slow down after peaking around 2040. This slowdown is attributed to the finite number of tasks that computer vision can automate, the increasing difficulty of replacing remaining human employees in specialized roles, and the limited number of humans in certain niche professions. Future research will extend to large language models (LLMs), which are expected to automate more complex tasks with high precision, further expanding the scope of AI's impact.
Beyond the Headlines
The study's focus on computer vision tasks, while insightful, represents only a segment of AI's potential. The broader implications of AI, particularly with the advent of LLMs, could introduce new economic sectors and tasks that are currently unforeseen. The research also highlights the ethical and practical challenges of automation. While AI can identify issues like tumors, the human element of diagnosis, communication, and empathy remains irreplaceable. This underscores a future where human and artificial intelligence collaborate, with AI handling data-intensive, repetitive tasks and humans focusing on complex decision-making, creative problem-solving, and interpersonal interactions. The potential for shared model training costs among firms within an industry could also foster a more collaborative approach to AI adoption, potentially mitigating some of the competitive disparities identified.
