Rethinking AI Communication: Beyond Next-Word Prediction

In discussions about artificial intelligence, particularly in the context of large language models (LLMs), a common sentiment arises: “These systems are merely advanced math algorithms that predict the next word in a sentence. They lack true intelligence.” While concerns about intelligence are valid, they overlook a fundamental aspect of how an advanced AI could interact with us in the future.

Let’s envision a world centuries into the future where artificial general intelligence (AGI) exists. If such an intelligence is digital and artificial, it will still need a means to communicate, much like we do. This prompts a crucial question: what forms of interaction can we expect from an AI if not through continuous streams of language or action requests?

Rather than expecting an AGI to always have a singular, guaranteed action to execute, we might consider the more realistic scenario of a continuous distribution of thoughts and potential actions it can undertake. After all, human communication is rarely about presenting one definitive answer; it often comprises a complex interplay of considerations and responses.

From my experience in machine learning—having developed neural networks and deeply engaged with their underlying principles—I recognize that while LLMs may seem simplistic mathematically, they reveal greater intricacies when it comes to functionality. At their core, artificial intelligences will always rely on algorithms and mathematical models to produce outputs. Therefore, it’s essential to ask skeptics of AI this pivotal question: What alternative methods of output would you deem acceptable for an artificial intelligence?

What forms of interaction would differentiate true AI from a mere “sophisticated autocomplete”? Despite the sophistication of any model, it inevitably needs to convey its insights or actions in some manner. Next-token prediction, despite its perceived limitations, appears to be a practical approach to fulfilling this need.

In the grand scheme of our technological evolution, it is vital to embrace the current capabilities of AI while considering the potential advancements that lie ahead. The dialogue surrounding AI’s role and its capacity for interaction should evolve, too, encouraging us to think deeper about what true intelligence and communication could look like in an increasingly automated world.