The field of artificial intelligence has long been driven by the assumption that increased computing power would inevitably lead to true intelligence. However, recent years have seen only marginal improvements, despite the immense resources devoted to scaling. The returns on investment are no longer justifying the costs, prompting a re-examination of the current approach.

The issue lies in the prioritization of aspects such as pattern matching and retrieval, which, although impressive, do not necessarily translate to improvements in reasoning, planning, or handling novel problems. These complex areas are more challenging to measure and fund, resulting in a lack of progress.

Moreover, the current state of AI research is hindered by a lack of honest benchmarking and a reliance on task-specific overfitting, leading to exaggerated claims of general intelligence. To truly advance AI research, a shift in focus is necessary, exploring alternatives such as modular architectures, compositional learning, and more rigorous benchmarking methods.

Although these approaches may be less lucrative and harder to fund, they hold the potential for genuine breakthroughs. As the AI community continues to invest heavily in scaling, it is likely that researchers will be forced to re-evaluate their methods within the next five years, considering alternative approaches like multimodal reasoning or world models to achieve meaningful progress.

Photo by Sergio Zhukov on Pexels
Photos provided by Pexels