In 2024, the explosive growth of AI models seems to have slowed down compared to recent years. Has it hit a bottleneck? Not quite. In my view, we’re entering a turning point in technological innovation because the full potential of AI models is far from being realized.
AI isn’t just another “app”; it’s a “robot” that can perceive, understand, and make decisions—what we call an “agent.” Let’s be clear: robots are not the same as humans. Human perception is limited to a small slice of the electromagnetic spectrum, but the real world is full of physical, chemical, and other data waiting to be digitized. AI, through multimodal models, can convert information that humans can’t directly perceive—like visuals, sound, temperature, and more—into data, uncovering the meaningful connections within. That’s why sectors like smart manufacturing, precision agriculture, and environmental monitoring are just at the starting line. The potential for AI in these verticals remains untapped; they represent the blue ocean for the future growth of AI. And as software and hardware technologies evolve, the dimensions of data that AI can process will only expand, opening up new possibilities for future applications.
And the opportunities don’t stop there. For instance, large language models like GPT excel in handling sequence data, especially language. This advantage often leads developers to focus on maximizing the accuracy of these models, shaping them into “experts.” However, this architecture is fundamentally tied to statistical rules, making it difficult for the models to break free from the constraints of their training data and exhibit genuine innovation. Even with mechanisms like temperature adjustment, it remains challenging for these models to generate “low-probability” content. In essence, the stronger a model’s ability to produce correct content, the more limited its capacity for innovation becomes—a paradox that isn’t easily resolved through simple optimization.
From a different perspective, rather than trying to force language models to be innovative, we could start from a more fundamental level and develop AI models specifically designed for creativity. Take, for example, diffusion models that excel at generating artwork; they demonstrate a strong capacity for innovation. The core mechanism here involves gradually adding noise and making multi-stage corrections throughout the output process. Whether this approach can be extended to other domains is worth further investigation. We might even consider moving beyond the constraints of human language altogether, using high-dimensional vector languages from a machine perspective to describe the world, seeking out the underlying connections in this new “language.” This could ultimately lead to the formation of a so-called “world model.”
For AI developers, there is still a vast frontier to explore, with many opportunities to leapfrog ahead. If we set our sights further into the future, we’ll realize that robots (intelligent agents) have yet to integrate into human society on a large scale. There is ample room for AI’s growth. The industry will continue to see survival of the fittest, as is the nature of competition. But what truly determines a company’s future is whether we can break free from conventional thinking and discover those opportunities for innovation!