The Natural Language Bus
Moving beyond the thin veneer on top of GPT
As a child fascinated by computers — a fascination I have clung on to ever since — I once sketched my vision of the perfect machine. Inside my crude drawing was a speaker — not for playing music, but for telling the computer’s internals what to do in response to people pushing the buttons. In my young mind, this was how computers should work: a universal bus consisting of verbal commands that would be heard and understood by the rest of the machine.
Decades later, we seem to have arrived at a similar vision, but perhaps not in the way we should have. The rise of Large Language Models has ushered in an era where natural language has become the de facto communication protocol for AI systems. Every interaction, every command, every operation flows through this natural language “bus” — a universal translator that turns human intent into computer action.
But is this really progress?
The history of computing is filled with specialized protocols and interfaces, each carefully crafted for its domain. SQL speaks the language of databases. OpenGL communicates in the vernacular of graphics. HTTP orchestrates the web’s complex dance of requests and responses. These protocols are not just ways to communicate, but also frameworks for thinking about problems and abstractions.
By replacing these specialized protocols with natural language, we’re not just changing how we talk to computers — we’re fundamentally altering how we think about computation. It’s like replacing a precise surgical instrument with a Swiss Army knife. Yes, the knife is more versatile, but would you want your surgeon using one?
The current trend of putting ChatGPT in every text box represents both overkill and oversimplification. It’s overkill because we’re using a sledgehammer to drive a nail — employing sophisticated language models for tasks that could be handled more efficiently (not to mention using less resources) by simpler and more specialized protocols and models. It’s oversimplification because we’re losing the rich, domain-specific abstractions that specialized protocols provide.
Consider a simple database query. In SQL, the statement SELECT * FROM users WHERE age > 18 is precise, unambiguous, and efficient. Converting this to “show me all users who are over 18” adds unnecessary complexity and ambiguity. The natural language version might be more accessible to beginners, but I would argue that it is actually less expressive and more prone to misinterpretation.
Moving beyond the ubiquitous chatbot or search interface for LLMs is indeed a step forward. Human-centered AI (HCAI) tools that call language models behind the scenes, through their own specialized interfaces, show promise. But too often, this remains a thin veneer — the bare minimum of layers atop ChatGPT, without the deep integration and specialization that truly powerful tools require.
This isn’t to say that natural language interfaces don’t have their place. They excel in poorly-defined problem spaces where the precision of specialized protocols might be premature or constraining. They’re valuable for exploration, for learning, or for situations where we’re still discovering what the right abstractions should be.
The future isn’t about finding a universal translator for human-computer interaction. Rather, I think it’s about developing the right languages for the right conversations.
