Rethinking AI: Beyond Simple Word Prediction

The ongoing debate about the nature and capability of artificial intelligence often centers around a fundamental question: Can an AI be more than just a predictor of the next word, or token, in a sequence? While some skeptics argue that large language models (LLMs) are merely sophisticated mathematical outputs, I believe there’s room for a more nuanced discussion.

Consider this scenario: envision a future, whether 200, 400, or even 1000 years from now, where artificial general intelligence (AGI) exists. If such intelligence is primarily digital, it must find a way to communicate and interact with the external world. The notion of how an AGI conveys information raises an intriguing point: must it do so exclusively through a continuous stream of language, and is it unreasonable to think it would generate a range of actions or words, rather than a single definitive response?

As someone with a background in machine learning—an experience enriched by both professional endeavors and personal projects—I’ve dabbled in neural networks and even coded the backpropagation algorithm from the ground up. My familiarity with the fundamentals of LLM architectures has led me to reflect on their underlying mathematics. It’s true that at their core, these models operate on mathematical principles, but that doesn’t mean they lack depth or potential.

When we refer to something as “artificial intelligence,” it’s vital to remember that its essence lies in math and algorithms. Every algorithm must yield some form of output to be functional. This leads me to pose a thought-provoking question to the critics: What criteria do you consider vital for an output method to classify as worthy of being called AI?

How should such an intelligence engage with us, avoiding the simplistic label of just a “fancy auto-complete”? Regardless of how advanced we make our models, they must ultimately generate an output, and when comparing methodologies, next-token prediction stands out as a practical approach.

In summary, while we continue to grapple with definitions and implications of AI technology, it may be more beneficial to embrace the notion that expressing complex ideas doesn’t necessarily negate the utility of existing methods. Instead, we can view them as evolving solutions, capable of adapting and expanding their capabilities alongside our understanding of intelligence in both artificial and natural forms.