Designing AI Interfaces

Chapter 1. Understanding Large Language Models and Systems

In the pre-smartphone era, mobile phone users faced a challenge: texting was slow. Phones were constrained to nine-key keypads, meaning that to text your friend “Can you print out MapQuest directions?” on your Nokia 3310, you’d need to tap 84 times.

In the late 1990s a new solution appeared: T9 (text on nine keys). Instead of pressing 2-2-2 to get the letter C, you pressed 2 once. The system would guess which of the three letters (A, B, or C) you actually wanted based on the word you were building.

For example, to type “hello,” you’d press 4-3-5-5-6, and T9 would figure out you meant H-E-L-L-O rather than G-D-J-J-M or any of the other possible combinations.

While this innovation sped up the average user’s typing, that’s not what this predictive text system was designed to solve. Also in the late 1990s, a US startup called Tegic Communications began working on input methods for people with motor impairments, who often struggled with traditional keyboards. The company’s engineers explored how to build efficient text entry using only eight tracked eye positions, dramatically reducing the physical effort needed to compose words. Tegic found the same principle could transform everyday mobile phones, and went on to repurpose its accessibility work to accommodate the growing demand for efficient mobile email and text messaging, resulting in one of the defining mobile input systems of the early 2000s.

Early Predictive Language ...