Rethinking AI: Beyond Simple Word Prediction

In the contemporary discourse surrounding artificial intelligence, particularly in the realm of language models like LLMs (Large Language Models), one prominent critique is their characterization as mere sophisticated algorithms that predict the next word or token in a sequence. While it’s true that these models can be described mathematically, this perspective might underestimate the potential and complexity of future AI systems.

Let’s envision a future—perhaps 200, 400, or even 1000 years from now—where we have developed a true Artificial General Intelligence (AGI). In this scenario, communication with the external environment is essential. The question then arises: how will such an AGI convey its insights or intentions? Is it unreasonable to think that an AGI could express itself through a continuous stream of language, rather than a single predetermined action or response?

It’s important to note that I have a background in machine learning, having engaged with neural networks and even coding the backpropagation algorithm from scratch during my learning journey. I acknowledge that the architecture of current language models is rooted in statistical analysis—essentially numbers and algorithms. However, this mathematical foundation doesn’t diminish the capacity for such models to serve as powerful communicative tools.

When we think about artificial intelligence, it’s imperative to recognize that all intelligent systems require some form of output to interact with humans and their environments. Thus, I pose this question to those skeptical of our current models: what form of output would truly qualify as “intelligent” in your view? How should an AI engage with us so that it transcends being merely a “fancy autocomplete”?

Regardless of the complexities involved, every sophisticated model must generate some form of output. In this context, next-word prediction may not be an optimal label for the extensive capabilities these systems may offer. Instead, it could be seen as a foundational approach to facilitating more nuanced interactions, laying the groundwork for a richer form of communication between humans and future iterations of AI.

As we venture further into the realm of artificial intelligence, it’s vital to keep the conversation open, dynamic, and forward-thinking. The future of AI holds tremendous possibilities, and understanding the nature of its outputs is a crucial step in fostering that evolution.