A few years ago I was watching “Some Reflections on Early History” by J.C.R. Licklider, the author of “Man-Computer Symbiosis” (please excuse the anachronisms, and read “Man” as “Human”). Licklider was one of the key players developing the concept of “networked computers” and the personal computer. He’s credited as one of the “fathers of the internet,” directly writing memos while working at the Pentagon. So as a military guy talking about an “Intergalactic Computer Network” where “anyone could access data and programs from anywhere,” he was seen as a loony by his coworkers, because back then computers were used for just calculating salaries and census data and missile trajectories.

In that boring talk, he describes that his biggest disappointment with personal computers, and people in general, was that they won’t take the time to learn to use a new device, and the keyboard is boring and slow. He thought that we should have better ways to type stuff, and that the QWERTY keyboard is an unhappy coincidence. That we could get “more bits of information out of every key press”; like how fast, how hard, or where you type; or maybe use different phalanges to input different keys, or have gauntlets with accelerometers in there to type faster.

Based on that, through 2020 and 2021, my pandemic project was using my computer only from the terminal. I started to use a Matrix server with a WhatsApp bridge, so I could chat through the terminal with friends and family, browsed the internet with lynx, and overall wanted to have the computer make the least amount of compute necessary, so that the bits flowing from the network, to my machine, to my screen, to my eyes stayed as close to 1:1 as I could get them.

The bandwidth from “brain-to-computer” (what I type in) shall be minimized, so I get the exact behavior I want out of as few keystrokes as possible. The bandwidth from “computer-to-brain” (what I read on the screen) was also to be minimized: no extra pixels, only the letters conveying the exact information that I requested.

Funny, that made me practice a lot of “how to maximize conciseness,” very useful for talking to a language model. They have a similar bandwidth problem: the quality of LLMs depends a lot on how much signal you can pack into a prompt. Too much context, too many pictures, anything over 200k tokens makes the LLM much dumber. Bandwidth can be measured, and the signal-to-noise ratio matters a lot.