Matthew Mandel

Today, an app’s functionality is usually exposed for manipulation via an application-specific programming interface. Such interfaces conveniently abstract underlying functionality for easy usage but (1) are accessed via code and (2) are different for different applications. Since LLMs solved NLP, it’s now possible for natural language to be a simple, universal interface for interacting with any application without code. I could have a general-purpose personal agent with all of my data book a flight by talking in English to a Delta bot that can translate an English-language directive to concrete actions. 

There are a few reasons to think we won’t immediately add natural language interfaces everywhere. First, calls to LLMs are still pretty expensive. They look less like the zero-marginal cost software we’ve become used to over the last 15 or so years and more like a physical good with real variable costs. 

Interacting in natural language also invites indeterminacy. Whereas programmatic instructions are executed programmatically, instructions in natural language will leave more degrees of freedom in how they’re carried out. That indeterminacy might just lead to less efficient processes as the bots stumble their way to the right answer, but it could also invite imprecision and ultimately incorrect outcomes. Jointly, these factors add up to a generality tax. We have the tools for universal interfaces, but they come at a cost.

The generality tax probably means that natural language stays a source of interoperability of last resort. There will usually be a more efficient way for computers to talk to each other. What’s exciting, though, are the interactions that wouldn’t happen but for natural language as a viable interface.

When will people be willing to pay such a generality tax? Fault-tolerant higher value transactions might make sense. The cost of booking a hotel or flight will swamp the cost of the tokens involved, and if the bots mess up, you can usually adjust your itinerary manually without too much cost or friction. And doing your bookings via bots communicating in natural language might unlock greater flexibility in negotiating rates or exploring alternatives that would be hard to specify ex ante in code.